34850173

DOI: 10.1111/j.1475-679X.2008.00299.x Journal of Accounting Research Vol. 46 No. 5 December 2008 Printed in U.S.A.
How Much New Information Is There in Earnings?

R AY B A L L A N D L A K S H M A N A N S H I VA K U M A R Received 12 March 2008; accepted 24 June 2008
ABSTRACT
We quantify the relative importance of earnings announcements in providing new information to the share market, using the R 2 in a regression of securities calendar-year returns on their four quarterly earnings-announcement window returns. The R 2 , which averages approximately 5% to 9%, measures the proportion of total information incorporated in share prices annually that is associated with earnings announcements. We conclude that the average quarterly announcement is associated with approximately 1% to 2% of total annual information, thus providing a modest but not overwhelming amount of incremental information to the market. The results are consistent with the view that the primary economic role of reported earnings is not to provide timely new information to the share market. By inference, that role lies elsewhere, for example, in settling debt and compensation contracts and in disciplining prior information,including more timely managerial disclosures of information originating in the rms accounting system. The relative informativeness of earnings announcements is a concave function of size. Increased information during earnings-announcement windows in recent years is due only in part to increased concurrent releases of management forecasts.
Graduate School of Business, University of Chicago; London Business School. We are grateful for the comments of William Beaver, Robert Bowen, John Core, George Foster, Richard Frankel, Thomas Lys, Darren Roulstone, Terry Shevlin, Douglas Skinner (the editor), Andy Van Buskirk, Joe Weber, an anonymous referee, and participants at the 2007 Nicholas Dopuch Conference at Washington University, at seminars at University of Pennsylvania, Stanford University, and University of Washington, and at a brown bag workshop at University of Chicago. Ball gratefully acknowledges nancial support from the University of Chicago, Graduate School of Business. 975
Copyright
C
, University of Chicago on behalf of the Institute of Professional Accounting, 2008
976
R. BALL AND L. SHIVAKUMAR
There is no evidence of abnormal information arrival in the weeks surrounding earnings announcements. Substantial information is released in management forecasts and in analyst forecast revisions prior (but not subsequent) to earnings announcements.
1. Introduction
We propose a simple but robust method of quantifying the relative importance of earnings announcements in providing new information to the share market. The method provides a measure of the proportion of the total information incorporated in share prices over a year that is associated with its four quarterly earnings announcements. The measure is the R 2 value obtained from a regression of securities calendar-year returns on their four earnings-announcement window returns. Our principal result is that the four quarterly earnings announcements collectively are associated with approximately 6% to 9%, on average, of the total information incorporated in share prices over the year. If extreme observations are deleted, this falls to approximately 5% to 6%. The average quarterly earnings announcement thus is associated with about 1% or 2% of the total annual information. We also report abnormal share trading volume at quarterly announcements of 0.25%, approximately, of the annual trading volume. These results imply that earnings announcements provide a modest but not overwhelming amount of incremental information. The comparatively low surprise content of earnings announcements is to be expected, for a variety of reasons. By its nature, accounting earnings is low frequency (quarterly), not discretionary (announced every quarter, independent of whether there is substantial new information to report), and primarily backward-looking. Other information, and hence revision in share price, is comparatively high frequency, frequently discretionary (released only when there is substantial information to report), and both forward-looking and backward-looking. While an inkling of the relatively low informativeness of earnings announcements can be gleaned from a close reading of prior studies including Ball and Brown [1968], Beaver [1968], Foster, Olsen, and Shevlin [1984], and Bernard and Thomas [1990]the literature does not provide a robust and precise measure of how important earnings are as a source of new information to the share market. Perhaps as a consequence, in our experience there are widely divergent beliefs on this important issue. Thus, our primary objective is to provide a more precise and robust measure of the amount of new information in earnings than has hitherto been available. We employ deliberate redundancy in the term new information, to stress our focus on the surprise content of public earnings announcements. We do not address the role of earnings over longer intervals. Nor do we address directly the role of rms accounting systems in producing timelier information released by managers (such as monthly sales gures or earnings forecasts) or by other parties (such as analysts and the press). While
NEW INFORMATION IN EARNINGS
977
it might initially seem tempting to widen the event window to capture the price reaction to prior earnings-related information, and in particularto encompass analysts and managers earnings forecasts, when taken to the limit this would require the event window to begin at the rms very inception, since all information other than changes in expected returns (discount rates) that causes share price revision relates to present and future earnings. Furthermore, when the return window is widened beyond the effective announcement period, the research question no longer is the amount of earnings surprise. Widening the window to the scal year or even beyond is appropriate when testing the role of nancial reporting in contractual settings, such as debt and management compensation, in which settlement is madeand hence outcomes are determinedwith only an annual frequency. Timeliness then is more appropriately measured from regressions of annual earnings on annual returns, as in Basu [1997]. In contrast, short windows are appropriate for measuring the amount of new information in public announcements. 1 Our research design has several desirable properties. First, it does not require estimates of an earnings surprise variable. The only earnings-related data required are announcement dates, not quantities. Consequently, the design does not require estimation of earnings expectations, using proxies such as lagged earnings or analysts consensus forecasts, and thus it is not affected by errors in measuring expectations. A second attractive feature of this design is that, by allowing the regression slopes to differ from unity, the event-window price reaction is allowed to spill over into correlated price changes outside the window. It thus is less susceptible to underestimating the informational role of earnings arising from any initial market underreaction to earnings. 2 In comparison with simply studying return volatility within the earnings-announcement window, the research design therefore does not assume market efciency. Indeed, the method provides a test of efciency, since slopes greater (less) than unity imply initial market under(over)reaction to earnings-related information, and slopes equal to unity are consistent with market efciency. The research design therefore provides a new measure of earnings mispricing. We report that all four slope coefcients from the regression of annual returns on the quarterly event-window returns exceed unity, consistent with postearnings announcement drift (Ball and Brown [1968], Foster, Olsen, and Shevlin [1984], Bernard and Thomas [1989]) and with price momentum generally (Jegadeesh and Titman [1993]), though not all coefcients
1 Butler, Kraft, and Weiss [2007, p.183] draw a similar distinction between intraperiod and long-horizon timeliness. 2 If the market initially underreacts to earnings information, volatility inside the event window decreases and volatility outside the window increases. Both effects bias downward the estimated relative importance of earnings information in research designs that address only the event-window abnormal price reaction. However, our approach does not incorporate revisions in stock prices after the end of the calendar year.
978
are signicantly different from unity. The average of the four slopes is approximately 1.07 to 1.15, depending on the return specication, implying that only a minor portion of the price adjustment during the earningsannouncement window spills over into the remainder of the calendar year. This is a considerably lower estimate than obtained under conventional research designs. Recent years exhibit a sharp increase in the proportion of annual information released during the earnings event windows. The increase occurs perhaps as far back as 2000, and is particularly acute in the last three years of our sample, 2004 to 2006. It could be due to increased nancial reporting quality subsequent to changes in accounting standards or to changed manager and auditor incentives after Enron/Sarbanes-Oxley. It could be due to a reduction in analyst forecast activity, Regulation Fair Disclosure (FD), chance, changed product or factor market conditions, or a combination of factors. It appears to be more pronounced for larger rms, and does not appear to be caused by a change in the sample composition of listed rms over time. An increase in management forecasts released concurrently with earnings (Anilowski, Feng, and Skinner [2007]) explains only part of the change. The analysis can be extended to measure the relative amount of information that is produced during periods adjacent to the earnings event windows. The results suggest that information arrival is slightly lower than normal both before and after earnings announcements, which is somewhat surprising. In the weeks prior to earnings announcements, one might expect information production by managers, analysts, and investors (e.g., Kim and Verrecchia [1997]), but we nd no evidence of unusual price activity. In the weeks following earnings announcements, one might expect analysts to revise their forecasts of future earnings, but the data suggest that revision consists of incorporating the new information in the announcement into the forecasts, rather than producing new information. Our regression design also can be used to quantify the relative informativeness of some other variables, such as dividends and forecasts. To investigate the absence of unusual price volatility in the weeks surrounding earnings announcements, we study the earliest forecast revision by an analyst after each of the four quarterly earnings announcements, and the last revision before. The average abnormal adjusted R 2 from annual regressions of calendar-year returns on returns at the four earliest forecast revisions is approximately zero. For the four latest forecast revisions it is 4.7%, and it increases substantially in recent years. In contrast with earnings announcements, management forecasts are discretionary and occur less frequently and thus, when made, each is associated with approximately 25% of quarterly return volatility. This research design is simple, but we believe it sheds light on the amount of new information conveyed in earnings and thusindirectlyon the economic role of accounting earnings. If the primary economic role of accounting earnings is not to provide new information to the share market in
979
a timely fashion, then it might lie elsewhere, for example, in the settlement of contracts (Watts and Zimmerman [1986]) and in the conrmation (and hence disciplining) of prior information (Gigler and Hemmer [1998], Ball [2001]), including information derived from managers private access to their rms accounting systems and released to the market in various ways. Section 2 reviews the literature and ideas that form the background to this study. Section 3 describes the research design and sample, and section 4 outlines the event-window return volatility results. Section 5 expands the analysis to measure the extent of information produced during adjacent nonannouncement windows, while section 6 analyses the information content of management forecasts and analysts earnings forecast revisions. Section 7 briey outlines the trading volume results. Section 8 offers conclusions and some thoughts concerning future research.
2. Background
Even though earnings announcements undoubtedly contain an element of surprise, there are valid reasons not to expect them to provide substantial new information to the share market. One reason not to expect substantial new information in earnings announcements is relative frequency. Earnings announcements are low frequency, occurring quarterly (in some countries, semiannually). In comparison, revisions in expectations, and hence in share prices, are high frequency, occurring relatively continuously. This difference is due largely to the myriad sources of more timely information available to investors, including aggregate, industry, and rm-specic information. Examples of aggregate-level information are statistical releases by the Federal Reserve, the Bureau of Labor Statistics, and other Government agencies; prices from capital, foreign exchange, and derivatives markets; and releases of survey data on macroeconomic expectations, consumer condence, and sentiment. Industry-level information includes industry association surveys and reports, sales data, prices, hotel occupancy rates, airline capacity utilization, etc. Firm-specic information includes changes in management, changes in strategies, mergers and acquisitions, restructurings, new products, product sales gures and forecasts, factor and product prices, labor negotiations, security analyst reports and forecasts, rating agency reports, and management forecasts. Market participants combine this information with their knowledge of the business to form expectations about performance. Given this plethora of more timely information, it is not surprising that the majority of rms report earnings that are closely in line with investors and analysts expectations. Indeed, managers have incentives to minimize earnings surprises, and hence to release information on factor and product prices, product sales, etc., in a more timely fashion than by awaiting its incorporation in the formal earnings release. A second reason to expect earnings announcements to be untimely is that they are not discretionary. Many disclosures are selected as a function of their informativeness. For example, managers can be expected to issue
980
earnings forecasts only when they have substantial private information to make public. 3 Consistent with this selection hypothesis, we show that management earnings forecasts are considerably less frequent than earnings announcements but, when made, they are considerably more informative. In comparison, there is essentially no discretionary selection involved in whether to report earnings: Firms are required to report every quarter, independent of whether there is substantial new information to report. A third reason to expect earnings announcements to be largely anticipated is that accounting income is based primarily on backward-looking information, such as past product sales and past production costs. Accounting recognition principles and practices generally are geared toward reporting actual outcomes, not toward incorporating the latest revisions in expectations of future outcomes. Share price is based on all available information, both backward-looking and forward-looking. An element of forward-looking expectations is introduced by accounting accruals (Dechow [1994]), including loss-recognition accruals (Basu [1997]). In addition, accounting earnings follows a random walklike process and hence contains information about future earnings. Nevertheless, accounting earnings is a substantially more backward-looking variable than share price, and than revision in share price. The implication is the well-documented result that prices lead earnings (e.g., Ball and Brown [1968], Beaver, Lambert, and Morse [1980], Easton, Harris, and Ohlson [1992], Kothari and Sloan [1992]), and hence that the surprise content of earnings is limited. These reasons lead us to the expectation that earnings announcements are unlikely to be a major source of timely new information. Nevertheless, it could be easy to gain the opposite impression from the nancial press, which devotes substantial coverage to announcements and to price changes around the day they are made. While some of the price changes can appear substantial, it is important to note that there is no baseline adjustment here for the underlying volatility of share prices, that is, for the typical magnitude of the price changes that occur in the absence of earnings announcements. Further, no adjustment is being made for selection effectsthat is, for the press focusing its reporting on the larger earnings and price changes. The fact that earnings announcements contain an element of surprise does not in itself attest to their relative importance as a source of new information. A similar impression could be garnered from much of the academic literature, which has a checkered history on this issue. Ball and Brown [1968, p. 176] draw two contrasting conclusions:
The initial objective was to assess the usefulness of existing accounting income numbers by examining their information content and timeliness. . . . Its content is . . . considerable. However, the annual income report does not rate highly as a timely medium.
3 Motives for informing the market might include reducing capital costs (Botosan [1997]), avoiding stockholder litigation (Skinner [1997]), trading on better terms as insiders (Penman [1982]), or complying with securities law.
981
Despite the clear conclusion that earnings are lacking in timeliness, Bamber, Christensen, and Gaver [2000] document subsequent literature leaving the impression that earnings announcements are periods when substantial new information is released. Those authors attribute this in part to a misinterpretation of Beaver [1968]. Bamber, Christensen, and Gaver [2000, p. 111] claim that the literature ignored the effect on Beavers [1968] results of an unrepresentative sample, consisting of 143 small non-12/31 rms with fewer than 20 annual Wall Street Journal news items. 4 They conclude that subsequent evidence of low timelinessreported, for example, in Oppong [1980], Atiase [1985], and Bamber [1986, 1987]did not substantially inuence the literatures interpretation of Beavers [1968] results. 5 We believe an additional problem in interpreting many prior studies arises from the way those studies explicitly or implicitly implement baseline adjustments. It can be misleading to compare event-window price behavior with equivalent behavior over nonevent weeks or days, without placing these short windows in a longer-term perspective. While these analyses indicate how earnings-announcement price behavior compares with typical nonannouncement days, they can give a misleading impression of how signicant the event-window price behavior is in the overall information environment. Thus, if price volatility on quarterly earnings-announcement days is double its typical daily level, a graphical representation of daily volatility will reveal a substantial spike at the eventyet the increase in volatility is only one-sixtieth of a quarters total volatility (assuming, for illustration, that nonannouncement daily returns are identically and independently distributed (i.i.d.) and 60 trading days per quarter). Similarly, Beaver [1968] reports a 67% increase in the squared unexpected price changes in the earnings-announcement week relative to nonannouncement weeks, and a 33% increase in trading volume. While graphs of these statistics in event time reveal a sharp increase at the announcement, the increases are on the order of only 5% and 2%, respectively, of the total quarterly price behavior (assuming Beavers [1968] results for annual earnings apply to every quarter, assuming that nonannouncement weekly returns are i.i.d., and ignoring any upward bias due to sample selection or skew). We propose that total price volatility over at least the scal period of the earnings report is a more appropriate baseline for assessing the relative importance of earnings announcements in providing new information to
4 As they take care to point out, the misinterpretation occurs in subsequent literature. Beaver [1968, p. 72] clearly outlines his sample selection criteria, and even cautions the reader as follows: As long as the criteria are visible ex ante, the population for which the studys ndings are relevant can be easily identied. Also, the sample criteria can be relaxed in future studies to discover the generality of the ndings presented here for other populations. 5 Ironically in view of their own hypothesisthat the literature overgeneralizes Beavers [1968] results from a small, unrepresentative sampleBamber, Christensen, and Gaver [2000] do not emphasize that Atiases [1985] sample also is small and unrepresentative. Atiase [1985] studies just 200 rms in two quarters (the second quarters of 1971 and 1972).
982
investors, because it benchmarks earnings relative to all information sources over the same interval. To be sure, a careful reading of Ball and Brown [1968], Beaver [1968], Morse [1981], Chambers and Penman [1984] Foster, Olsen, and Shevlin [1984], Bernard and Thomas [1990], Collins, Kothari, Shanken, and Sloan [1994] Beaver [1998, p. 105106], Nichols and Wahlen [2004], and other studies could support the conclusion that earnings announcements are not a timely source of information, even though only Ball and Brown [1968] explicitly state that conclusion. In addition, the results of Oppong [1980] and Atiase [1985] could cause strong doubts as to the extent of earnings informativeness. Similarly, to highlight the effects of small rms in Beaver [1968], Bamber, Christensen, and Gaver [2000] show that the magnitude of the market reaction at earnings announcements by Fortune 200 rms, which constitute two-thirds of the New York Stock Exchanges total market value, is statistically indistinguishable from price movements in several of the weeks surrounding the earnings announcements. Recently, Beekes and Brown [2006a,b] reformulate the Ball and Brown [1968] analysis of annual earnings timeliness. Further, Butler, Kraft, and Weiss [2007] report that, prior to the introduction of mandatory quarterly reporting, there was little difference between the price behavior of rms reporting quarterly and those reporting only semiannually, suggesting that a higher frequency of reporting does not add substantial new information. There thus is available evidence to doubt that earnings are a source of substantial new information. No study to our knowledge has quantied the importance of earnings as a timely source of information relative to the total information environment. Perhaps as a consequence, in our experience there are widely divergent opinions on this issue. The method we use to estimate the relative informativeness of earnings announcements measures the proportion of the total information incorporated in share prices over a year that is associated with the four quarterly earnings announcements during the year.
3. Research Design and Sample

To provide a more precise and more readily interpretable measure of how much value-relevant information is provided at earnings releases, we calculate the adjusted R 2 from a regression of calendar-year returns on returns in each of the four three-day announcement event windows during the year. We estimate annual cross-sectional regressions of calendar-year returns (R i (annual)) on returns in the four quarterly earnings-announcement windows (R i (windowK ), K = 1 to 4) during the calendar year: Ri (annual) = a0 + a1 Ri (window1) + a2 Ri (window2) + a3 Ri (window3) + a4 Ri (window4) + i (1)
The regression adjusted R 2 measures the proportion of annual return variability associated with the four earnings event windows.
983
We estimate the regression from annual cross-sections. Similar results are obtained from rm-level time-series regressions, though they are more likely to be affected by survivorship bias. In the following subsections, we discuss issues relating to empirical implementation and interpretation.
3.1
SAMPLE
The sample comprises all rm-years with data available on the quarterly Compustat and daily Center for Research in Security Prices (CRSP) return les from January 1972 to December 2006. The sample period is determined by the availability of earnings-announcement dates on Compustat. The earnings-announcement window is dened as day 1 to day +1, where day 0 is the Compustat announcement date, though our research design is relatively insensitive to the length of the window chosen. We include only rm-years that have exactly four earnings announcements in the calendar year, whose four event windows lie entirely within the calendar year, and that have returns data on the CRSP daily le for at least 240 trading days (out of approximately 252 trading days, on average). The last of these requirements eliminates an average of 25 rms per year (0.7% of the observations).
3.2
RETURN MEASUREMENT
Calendar-year buy-and-hold returns are computed from the daily returns. Earnings-announcement returns are buy-and-hold returns over days 1 to +1, relative to the Compustat earnings-announcement date, day 0. The convention followed is that the quarter designates the calendar quarter in which the earnings announcement occurs, which typically is in the quarter following the scal period over which earnings are calculated. For example, quarter 1 is January through March, which for a December 31 rm is the calendar quarter in which the prior scal years and fourth quarters earnings typically are released. The results are somewhat sensitive to whether returns are calculated in arithmetic or logarithmic terms, so both are presented. Log returns can be meaningfully aggregated across time, and hence are more consistent with a linear regression of annual returns on event-window returns. However, averaging logarithmic returns across a portfolio of stocks is problematic, and they exhibit some left skew. Conversely, arithmetic returns can be logically aggregated across securities, but not time, and exhibit considerable right skew. The appropriate choice is unclear, so we present both.
3.3
BENCHMARKING THE REGRESSION
R2
To infer the information content of earnings announcements, we compare the adjusted R 2 value from regression (1) against a benchmark that is designed to reect normal price volatility. Specically, we calculate the abnormal R 2 value as the difference between the regression adjusted R 2 value and its expected value under the null hypothesis that daily returns, including earnings-announcement window returns, are i.i.d. across time. In an average year with 252 trading days, the four three-day windows then would
984
contain approximately 4.8% (12/252) of the total annual information, so we take 4.8% to be the expected value of the adjusted R 2 under the null. 6 Evidence of serial correlation in daily returns challenges the validity of the i.i.d. assumption underlying this benchmark. For instance, French and Roll [1986] show that daily returns are negatively autocorrelated due to bidask bounce in prices. It therefore might seem that a more appropriate benchmark would be obtained from regressions of annual returns on randomly chosen three-day windows that do not include earnings announcements. This procedure can be expected to provide a lower benchmark than obtained when ignoring the negative serial correlation. However, Bernard and Thomas [1989] document positive autocorrelation in daily returns conditional on earnings announcements, which implies that the benchmark might be adjusted upward. The issue is whether it is appropriate to force the negative correlation in non-earnings-announcement windows onto earnings-announcement windows, where such correlation does not exist and if anything is reversed. To examine how serial correlation affects the benchmark, we re-estimate regression (1) for the primary sample in tables 2 and 3 below, but replacing the four actual event-window returns each rm/year with returns for the same rm/year in four randomly chosen three-day windows. A random date is chosen from the 13 weeks surrounding each earnings-announcement date, subject to the requirement that the date is in the same calendar year and on the same day of the week as the actual earnings-announcement date. The regressions are otherwise identical to those reported in tables 2 and 3. The mean adjusted R 2 value in these regressions is 3.8% for arithmetic returns and 5.3% for logarithmic returns. This is 1.0% lower than our 4.8% annual benchmark for arithmetic returns (0.25% per quarter lower) and 0.5% higher than our benchmark for logarithmic returns (0.125% per quarter). These simulations do not incorporate the observed positive autocorrelation in returns conditional on earnings announcements, and hence understate the adjusted R 2 values. While the precise appropriate benchmark is debatable, the simulation results suggest that the benchmark of 4.8% under the i.i.d assumption is not unreasonable, and that the choice of benchmark does not affect the tenor of our conclusions.
4. Event-Window Contribution to Annual Return Volatility

4.1
SUMMARY STATISTICS
Table 1 reports annual summary statistics for calendar-year returns and for returns in each of the four quarterly earnings-announcement windows.
6 The expected R 2 value is computed annually, based on the number of trading days in that year, which varies between 248 in 2001 and 254 in 1992 and 1996. The expectation therefore varies only slightly, between 4.7% and 4.8%.
985
TABLE 1 Summary Statistics for Returns in Calendar Years and in the Four Earnings-Announcement Windows Panel A: Arithmetic returns No. of % Obs. % Obs. Obs. Mean Median Skewness = 0 >0 Calendar-year 19722006 mean 3,682 0.182 0.089 4.065 0.36 59.1 returns t-statistic 4.47 2.71 5.76 19721989 mean 2,492 0.166 0.103 2.250 0.39 61.3 19902006 mean 4,941 0.199 0.075 5.987 0.33 56.8 Earnings-announcement 19722006 mean window returns in t-statistic quarter 1 19721989 mean 19902006 mean Earnings-announcement 19722006 mean window returns in t-statistic quarter 2 19721989 mean 19902006 mean Earnings-announcement 19722006 mean window returns in t-statistic quarter 3 19721989 mean 19902006 mean Earnings-announcement 19722006 mean window returns in t-statistic quarter 4 19721989 mean 19902006 mean Panel B: Logarithmic returns No. of Obs. 19722006 mean 3,682 t-statistic 19721989 mean 2,492 19902006 mean 4,941 3,682 2,492 4,941 3,682 2,492 4,941 3,682 2,492 4,941 3,682 2,492 4,941 Mean 0.034 0.89 0.062 0.005 0.003 2.43 0.005 0.001 0.003 2.25 0.002 0.003 0.002 1.36 0.000 0.004 0.001 0.59 0.003 0.001 Median Skewness 0.069 0.466 2.13 4.15 0.077 0.320 0.060 0.620 0.001 2.09 0.001 0.001 0.001 2.00 0.001 0.002 0.000 0.04 0.001 0.001 0.000 0.29 0.002 0.002 0.167 1.59 0.300 0.027 0.281 1.90 0.348 0.210 0.327 1.98 0.070 0.747 0.120 1.03 0.134 0.389 % Obs. % Obs. =0 >0 0.36 59.1 0.39 0.33 2.13 2.81 1.41 2.36 3.13 1.53 2.36 3.24 1.43 2.38 3.31 1.40 61.3 56.8 50.6 50.7 50.4 50.2 49.0 51.4 48.4 48.1 48.7 48.5 46.7 50.5 3,682 2,492 4,941 3,682 2,492 4,941 3,682 2,492 4,941 3,682 2,492 4,941 0.006 4.97 0.007 0.005 0.006 4.28 0.004 0.008 0.001 1.23 0.002 0.001 0.003 1.75 0.000 0.006 0.001 2.09 0.001 0.001 0.001 2.01 0.001 0.002 0.000 0.05 0.001 0.001 0.000 0.27 0.002 0.002 1.754 7.76 1.213 2.327 2.901 3.14 1.193 4.710 1.192 5.44 1.032 1.362 1.975 5.61 1.690 2.277 2.13 2.81 1.41 2.36 3.13 1.53 2.36 3.24 1.43 2.38 3.31 1.40 50.6 50.7 50.4 50.2 49.0 51.4 48.4 48.1 48.7 48.5 46.7 50.5
Calendar-year returns
Earnings-announcement 19722006 mean window returns in t-statistic quarter 1 19721989 mean 19902006 mean Earnings-announcement 19722006 mean window returns in t-statistic quarter 2 19721989 mean 19902006 mean Earnings-announcement 19722006 mean window returns in t-statistic quarter 3 19721989 mean 19902006 mean Earnings-announcement 19722006 mean window returns in t-statistic quarter 4 19721989 mean 19902006 mean
This table presents summary statistics for calendar-year returns and for earnings-announcement returns. Calendar-year buy-and-hold returns are computed from daily CRSP returns. Earnings-announcement returns are buy-and-hold returns for the three days surrounding the Compustat announcement date. Firm-years with other than four earnings announcements or with daily return data for fewer than 240 trading days are excluded. Annual statistics (mean, median, skew) from the distribution of return across stocks are computed for the calendar-year return and for the four earnings-announcement window returns. The table presents means of these statistics across years, with associated FamaMacbeth t-statistics. Panel A presents statistics for arithmetic returns, while panel B presents results for logarithmic returns.
986
Panels A and B report statistics for arithmetic and logarithmic returns, respectively, calculated from identical rm-year samples. The different distributional properties of arithmetic and logarithmic returns are apparent. In the average calendar year over 1972 to 2006, the median arithmetic and logarithmic returns are similar (8.9% and 6.9%, respectively), but the means (18.2% and 3.4%, respectively) and the skew statistics (+4.065 and 0.466, respectively) differ substantially. The mean 1972 to 2006 logarithmic return (3.4%) is not even signicant (FamaMacBeth t-statistic of 0.89). Analogous differences are observed in each of the four three-day event-window return distributions, and in both the 1972 to 1989 and 1990 to 2006 subperiods. The rst two quarterly earnings-announcement windows in the calendar year experience signicant positive mean and median arithmetic and logarithmic returns. The arithmetic (logarithmic) mean announcement window return for both quarters is 0.6% (0.3%) over three days, which is approximately 50% (25%) on an annualized basis. This is consistent with an expected risk premium during earnings-announcement periods. 7 However, the medians during the announcement windows are only 0.1%. Both the mean and median earnings-announcement returns in the third and fourth quarters are not signicantly different from zero. The p-value from an F test for equal mean returns in the four announcement windows is 0.04 for arithmetic returns and 0.02 for logarithmic returns. In every calendar quarter, more than 2% of all rms experience zero returns in the three-day earnings windows, consistent with an absence of material new information for these rms. The frequency of zero event-window returns falls substantially over time.
4.2
ABNORMAL
R 2 : NEW INFORMATION IN EARNINGS
Tables 2 and 3 report results for arithmetic and logarithmic returns, respectively, calculated from identical rm-year samples. The sample size rises from 1,405 rms in 1972 to 6,140 rms in 2000, and falls to 4,822 rms in 2006. Panel A presents regression estimates for individual years, and panels B through D summarize them for the full sample, December year-end rms, and non-December year-end rms, respectively. For the full sample over 1972 to 2006 in panel A, the average abnormal adjusted R 2 from the annual regressions is 5.9% for arithmetic returns and 9.3% for logarithmic returns, allowing for the 4.8% expected value under the null. Individual quarterly earnings announcements therefore are associated, on average, with approximately 1.5% to 2.3% of all the value-relevant information affecting prices in a year. The abnormal R 2 is positive in 34
7 See Chari, Jagannathan, and Ofer [1988], Ball and Kothari [1991], Lamont [1998], Cohen et al. [2007], and Lamont and Frazzini [2007]. Consistent with this literature, untabulated results show that the mean announcement-window returns are substantially larger for smaller stocks.

TABLE 2 Annual Regressions of Calendar-Year Returns on the Four Announcement Window Returns (Arithmetic Returns) Year Panel A: All rms 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 Intercept 0.067 0.319 0.264 0.589 0.423 0.136 0.141 0.373 0.302 0.048 0.291 0.332 0.054 0.289 0.071 0.087 0.219 0.165 0.169 0.462 0.184 0.221 0.023 0.338 0.169 0.224 0.041 0.384 0.056 0.193 0.116 0.738 0.223 0.061 0.171 Quarter 1 0.758 0.516 0.793 0.982 0.877 0.971 1.506 0.894 0.703 1.294 1.712 1.379 1.025 1.050 1.087 1.152 1.085 1.277 0.484 1.700 3.402 1.299 0.778 1.556 0.774 1.248 0.974 1.370 0.695 1.238 1.268 2.049 1.473 1.245 1.256 Quarter 2 1.160 0.666 0.943 1.397 1.136 0.874 1.187 1.457 0.970 1.048 1.054 0.989 0.856 1.651 1.147 1.061 1.195 1.360 0.729 2.568 0.832 1.395 0.887 0.764 1.565 0.945 0.828 1.145 0.905 0.837 1.317 2.525 1.220 1.319 1.020 Quarter 3 1.396 0.602 0.522 1.660 1.444 1.556 0.636 1.013 1.490 1.080 1.274 1.425 0.602 1.141 0.900 1.069 1.593 1.306 0.887 1.083 0.808 1.427 0.680 1.324 1.194 0.880 0.899 1.549 1.377 1.536 0.547 1.341 1.053 1.473 1.041 Quarter 4 0.780 0.835 0.888 2.057 1.128 0.897 0.419 1.843 1.272 1.318 1.829 1.018 1.120 0.880 0.991 0.072 1.178 1.004 0.554 1.287 1.297 1.345 1.174 1.406 1.177 1.149 0.576 1.999 0.730 0.659 0.262 2.607 1.007 1.130 1.064 Abnormal R2 % 7.4 3.8 7.1 8.2 5.5 4.9 3.9 1.9 2.2 11.4 7.7 6.8 6.1 3.8 7.9 5.6 7.6 6.5 2.9 7.9 7.3 6.7 4.5 1.9 4.5 4.8 1.1 1.1 5.1 1.6 8.8 6.6 9.7 13.9 11.7
987
No. of Obs. 1,405 1,778 2,211 2,273 2,283 2,316 2,282 2,202 2,187 2,146 2,106 2,634 3,012 3,091 3,011 3,265 3,277 3,385 3,393 3,515 3,772 4,076 4,537 5,042 5,395 5,644 5,840 5,528 6,140 5,868 5,484 5,100 4,953 4,890 4,822
Panel B: Time-series summary statistics (full sample) 19722006 mean 0.162 1.196 1.170 1.137 p-value (H 0 = 1) 0.03 0.02 0.02 19722006 median p-value (H 0 = 1) 19721989 mean 19902006 mean 0.171 0.151 0.174 1.152 0.00 1.059 1.342 1.061 0.02 1.119 1.224 1.141 0.00 1.150 1.123
1.113 0.20 1.120 0.00 1.085 1.143
5.9 6.1 6.0 5.8
3,682 3,385 2,492 4,941

(Continued)
988

T A B L E 2 Continued Intercept Quarter 1 Quarter 2 Quarter 3 Quarter 4 Abnormal R2 % 5.5 5.4 5.5 5.5 No. Obs. 2,368 1,923 1,484 3,305
Panel C: Time-series summary statistics (December scal year-end sample) 19722006 mean 0.161 1.102 1.161 1.176 1.143 p-value (H 0 = 1) 0.10 0.04 0.02 0.16 19722006 median 0.172 1.177 1.081 1.138 1.142 p-value (H 0 = 1) 0.00 0.00 0.00 0.00 19721989 mean 0.152 0.986 1.126 1.197 1.111 19902006 mean 0.170 1.226 1.197 1.153 1.176
Panel D: Time-series summary statistics (non-December scal year-end sample) 19722006 mean 0.164 1.268 1.190 1.067 1.074 6.9 p-value (H 0 = 1) 0.05 0.04 0.31 0.40 19722006 median 0.177 1.194 1.067 1.076 1.046 6.6 0.00 0.10 0.02 0.74 p-value (H 0 = 1) 19721989 mean 0.150 1.135 1.137 1.063 1.055 7.0 19902006 mean 0.178 1.408 1.246 1.070 1.094 6.7
1,307 1,368 1,005 1,628
This table presents estimates from the following regression of calendar-year arithmetic returns (R i (annual)) on the four quarterly earnings-announcement returns in the year (R i (window1) to R i (window4)): Ri (annual) = a0 + a1 Ri (window1) + a2 Ri (window2) + a3 Ri (window3) + a4 Ri (window4) + i Calendar-year arithmetic buy-and-hold returns are computed from daily CRSP returns. Earningsannouncement returns are arithmetic buy-and-hold returns for the three days surrounding the Compustat announcement date. Firm-years with other than four earnings announcements or with daily return data available for fewer than 240 trading days are excluded. Panel A presents annual regression coefcients and abnormal R 2 values for all rms. Panels B to D present time-series summary statistics for the full sample and December and non-December scal year-end rms. The p-values are from F -tests that the mean coefcient is 1.0 and sign tests that the frequency of coefcients exceeding 1.0 is 0.5. Abnormal R 2 is the regression adjusted R 2 value minus its expectation assuming i.i.d. daily returns.
of the 35 years (arithmetic) and in all years (logarithmic), and hence is statistically signicant under a binomial sign test at a p-value < 0.0001. Three points suggest these could be upward-biased estimates. First, the R 2 values reect the price response to all information that is released during the earnings-announcement window, not only earnings. Concurrent information includes management commentary on the nancials, management earnings forecasts (we report evidence on them below), press commentary, dividend announcements, restructuring plan announcements, non-GAAP street earnings, and analyst forecast revisions. 8 Second, because the average regression slope coefcients exceed unity, the 1.5% to 2.3% estimate incorporates some correlated price response that occurs outside the threeday announcement period. Third, in the following subsection we report that the abnormal R 2 values are substantially reduced when outliers are excluded. Overall, the result that quarterly earnings announcements are associated with 1.5% to 2.3% of the total annual price volatility implies they contain a modest but not overwhelming amount of new information.
8
Francis, Schipper, and Vincent [2002] and Collins, Li, and Xie [2008].

TABLE 3 Annual Regressions of Calendar-Year Returns on the Four Announcement Window Returns (Logarithmic Returns) Year Panel A: All rms 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 Intercept 0.017 0.491 0.383 0.409 0.315 0.087 0.093 0.254 0.207 0.005 0.183 0.239 0.119 0.190 0.005 0.174 0.134 0.066 0.308 0.255 0.070 0.100 0.116 0.172 0.050 0.086 0.211 0.047 0.359 0.026 0.300 0.436 0.131 0.016 0.100 Quarter 1 0.844 1.131 0.945 0.570 0.663 0.838 1.116 0.806 0.701 1.287 1.401 1.135 1.403 0.932 1.101 1.305 0.966 1.109 0.581 0.967 1.231 1.117 1.041 1.064 0.955 1.227 1.046 0.837 1.154 1.012 1.726 0.883 1.260 1.303 1.103 Quarter 2 1.202 1.066 1.334 0.860 0.886 0.835 0.982 1.041 0.728 1.084 1.020 0.727 1.165 1.396 1.273 1.319 1.070 1.337 1.140 1.111 0.915 1.039 0.942 1.077 1.143 1.068 1.004 0.828 1.340 0.758 1.449 1.112 0.998 1.155 0.964 Quarter 3 1.367 1.202 0.926 0.925 0.982 1.301 0.679 0.710 1.121 1.135 1.021 1.103 0.801 0.980 0.999 1.378 1.483 1.340 1.459 0.933 0.942 1.169 0.839 1.124 0.978 0.988 1.066 1.087 2.268 1.402 0.960 0.708 0.889 1.563 0.998 Quarter 4 0.963 1.315 1.054 1.189 0.859 0.735 0.354 1.212 0.908 1.283 1.488 0.972 1.588 0.904 1.128 0.317 1.112 1.220 1.043 1.045 0.900 0.999 1.251 1.130 1.256 1.351 0.907 0.977 1.602 0.705 0.525 0.931 0.852 1.085 1.095 Abnormal R2 % 11.7 10.0 12.0 9.0 8.9 6.7 5.5 4.6 4.5 12.3 9.9 9.2 9.4 6.4 9.2 7.0 10.5 8.9 8.3 8.0 10.5 10.4 7.5 8.2 8.5 9.3 5.5 2.9 11.5 6.6 10.3 8.7 15.9 20.4 15.9
989
No. of Obs. 1,405 1,778 2,211 2,273 2,283 2,316 2,282 2,202 2,187 2,146 2,106 2,634 3,012 3,091 3,011 3,265 3,277 3,385 3,393 3,515 3,772 4,076 4,537 5,042 5,395 5,644 5,840 5,528 6,140 5,868 5,484 5,100 4,953 4,890 4,822
Panel B: Time-series summary statistics (full sample) 19722006 mean 0.033 1.050 1.068 1.109 p-value (H 0 = 1) 0.23 0.04 0.04 19722006 median p-value (H 0 = 1) 19721989 mean 19902006 mean 0.070 0.057 0.006 1.064 0.02 1.014 1.089 1.068 0.00 1.074 1.061 1.021 0.74 1.081 1.140
1.036 0.47 1.045 0.32 1.033 1.038
9.3 9.0 8.6 9.9
3,682 3,385 2,492 4,941

(Continued)
990

T A B L E 3 Continued Intercept Quarter 1 Quarter 2 Quarter 3 Quarter 4 Abnormal R2 % 8.6 8.2 7.5 9.8 No. Obs. 2,368 1,923 1,484 3,305
Panel C: Time-series summary statistics (December scal year-end sample) 19722006 mean 0.039 1.022 1.037 1.134 1.072 p-value (H 0 = 1) 0.63 0.26 0.03 0.28 19722006 median 0.074 1.041 1.040 1.093 1.099 p-value (H 0 = 1) 0.02 0.10 0.00 0.10 19721989 mean 0.068 0.939 1.017 1.082 1.072 19902006 mean 0.009 1.111 1.058 1.189 1.071
Panel D: Time-series summary statistics (non-December scal year-end sample) 19722006 mean 0.023 1.068 1.099 1.055 0.989 10.4 p-value (H 0 = 1) 0.14 0.04 0.29 0.81 19722006 median 0.068 1.044 1.125 1.019 0.988 10.1 0.02 0.00 0.32 0.74 p-value (H 0 = 1) 19721989 mean 0.042 1.089 1.143 1.047 0.994 10.4 19902006 mean 0.003 1.046 1.052 1.064 0.983 10.3
1,307 1,368 1,005 1,628
This table presents estimates from the following regression of calendar-year logarithmic returns (R i (annual)) on the four quarterly earnings-announcement returns in the year (R i (window1) to R i (window4)): Ri (annual) = a0 + a1 Ri (window1) + a2 Ri (window2) + a3 Ri (window3) + a4 Ri (window4) + i Calendar-year logarithmic buy-and-hold returns are computed from daily CRSP returns. Earningsannouncement returns are logarithmic buy-and-hold returns for the three days surrounding the Compustat announcement date. Firm-years with other than four earnings announcements or with daily return data available for fewer than 240 trading days are excluded. Panel A presents annual regression coefcients and abnormal R 2 values for all rms. Panels B to D present time-series summary statistics for the full sample and December and non-December scal year-end rms. The p-values are from F -tests that the mean coefcient is 1.0 and sign tests that the frequency of coefcients exceeding 1.0 is 0.5. Abnormal R 2 is the regression adjusted R 2 value minus its expectation assuming i.i.d. daily returns.
4.3
DELETING EXTREME OBSERVATIONS
To assess whether results are sensitive to extreme outliers, we re-estimate them for a trimmed sample that excludes, in each calendar year, the extreme 1% on either side of the calendar-year and announcement-window returns. Deletion of an individual quarter requires deletion of the entire rm/year. In untabulated results, the average annual abnormal R 2 values fall substantially, for the entire period and also for the two subperiods. The full-period average falls from 5.9% to 4.8% (arithmetic) and from 9.3% to 6.0% (logarithmic). Thus, after deleting outliers, individual quarterly earnings announcements are associated with 1.2% to 1.5% of the total annual price volatility.
4.4
DECEMBER FISCAL YEAR-END FIRMS
The abnormal R 2 values, on average, are approximately one-fth larger for non-December scal year-end rms than for December scal year-end rms. For the arithmetic returns in table 2, the averages are 5.5% for December scal year-end rms and 6.9% for non-December scal year-end rms. For the logarithmic returns in table 3, the respective averages are 8.6% and 10.4%. These results are consistent with prior evidence that the earnings of December scal year-end rms are less incrementally informative (Bamber, Christensen, and Gaver [2000]).
991
4.5
GOOD AND BAD AGGREGATE NEWS
We next examine how the average R 2 value varies in years with predominantly good news versus those with predominantly bad news, dened as years with positive and negative aggregate sample mean returns, respectively. In untabulated results, the 11 (24) calendar years where the mean returns are negative (positive) exhibit an average abnormal adjusted R 2 value of 5.0% (6.2%), using arithmetic returns. For logarithmic returns, the corresponding averages are 9.1% and 9.3%.
4.6
SLOPE COEFFICIENTS: MISPRICING EARNINGS
For the full sample in panel A and for all years 1972 to 2006, and for both arithmetic and logarithmic returns, each of the four slope coefcients from the regression of annual returns on the quarterly event-window returns exceeds unity. The coefcients are statistically signicantly different from one at the conventional 5% level for event windows in the rst through third calendar quarters for arithmetic returns, and in the second and third calendar quarters for logarithmic returns. This result is consistent with the seemingly delayed reaction to earnings known as postearnings announcement drift (e.g., Ball and Brown [1968], Bernard and Thomas [1989, 1990]). It also is consistent with delayed reaction to information in periods prior to the earnings announcement that is corrected at the time of the announcement, and with price momentum generally (Jegadeesh and Titman [1993]). Economically, the slopes are close to unity, the expected coefcient if the market correctly prices earnings and earnings-related information (specically, if there is no systematic under- or overreaction that is corrected within the same calendar year). The average of the four slopes is approximately 1.15 for arithmetic returns (table 2, panel B) and 1.07 for logarithmic returns (table 3, panel B), implying that only 0.005 (0.072 ) to 0.025 (0.152 ) of the price volatility during the three-day earnings-announcement window spills over into volatility at other times during the calendar year. This is a considerably lower estimate of the proportion of delayed market reaction to earnings than is obtained by correlating postannouncement returns with earnings (Ball and Brown [1968], Foster, Olsen, and Shevlin [1984], Bernard and Thomas [1989]), as distinct from our design, which correlates them with earnings-announcement returns. We suspect that this occurs in part because our measure is relatively free of the bias that arises from earnings being correlated with expected returns and changes in expected returns (Ball [1978]). It also could be due to postannouncement returns in this research design being truncated at the end of the calendar year, which could particularly affect earnings announcements in the fourth quarter. Against this, we note that the design incorporates what could be labeled preearnings announcement driftunderreaction to earningsrelated information prior to the earnings announcement that is corrected at the time of the announcement. Trimming the data, by deleting 1% of the observations on both extremes of calendar-year and announcement-window returns, leads to estimated
992
FIG. 1.Abnormal adjusted R 2 values and slope coefcients from annual 1972 to 2006 cross-sectional regressions of calendar-year returns on returns at the four quarterly earnings announcements. Calendar-year buy-and-hold returns are computed from daily CRSP returns. Earnings-announcement returns are buy-and-hold returns for the three days surrounding the Compustat announcement date. The sample is all rm-years with available data on the quarterly Compustat and daily CRSP les. Firm-years with other than four earnings announcements or with daily returns data for fewer than 240 trading days are excluded.
slopes that are closer to (and not signicantly different from) the null of unity. These results are not tabulated.
4.7
INDIVIDUAL YEAR AND SUBPERIOD RESULTS
Figure 1 depicts the time series of annual abnormal R 2 values. There is a sharp increase during recent years in the proportion of annual price revision that occurs in the four earnings-event windows. In panel A of tables 2 and 3, the annual abnormal R 2 value increases sharply somewhere in the 2000 to
993
2006 period. The abnormal R 2 values in the last three years with data, 2004 to 2006, are the three highest of the 35 years studied (logarithmic returns), or three of the four highest (arithmetic returns). 9 A similar increase is observed in event-window abnormal volume, expressed as a percentage of annual volume. The highest median abnormal volume occurs in the last four years, 2003 to 2006 (individual years are not reported in table 11 below). One explanation for the sharp increase is an increased relative informativeness of earnings in recent years. An increase in relative informativeness of earnings could arise from a real increase in exogenous earnings shocks, due, for example, to more volatile product and factor market conditions. It could arise from changes in accounting standards such as Statement of Financial Accounting (SFAS) 142 and SFAS 144, which set new rules for the impairment of intangible and tangible long-term assets, effective 2002, or from the Financial Accounting Standards Boards increased emphasis on fair value accounting generally. It could be due to changes in managers and auditors incentives, in response to the Enron era scandals or the enactment of the Sarbanes-Oxley Act of 2002. These events could have increased the monitoring of nancial reporting by parties such as auditors, internal auditors, boards, and the press, increasing the probability of detecting low-quality nancial reporting. At the same time, these events most likely increased the expected legal and political costs of being discovered to have engaged in low-quality reporting. An alternative explanation is deterioration over time in forecast accuracy, causing an increased surprise content of earnings. This could arise from a reduction in analyst forecast activity, due in turn to Regulation FD, which became effective in October 2000. It could arise from restrictions imposed on analysts subsequent to the biased-research scandals in 2000 to 2003, though a recent increase in the price reaction to analyst forecast revisions, reported in section 6, appears to rule this explanation out. 10 One feasible explanation, a change in sample composition, is made plausible by the fall in sample size from 6,140 to 4,822 rms between 2000 and 2006, and by the seemingly low abnormal R 2 values during the internet boom period. We investigate this by studying a constant sample of 1,593 rms with complete data in every year over 1996 to 2006. Results for arithmetic and logarithmic returns are presented in panels A and B of table 4, respectively. In both panels, the increase in the abnormal R 2 value over 2000 to 2006 is even sharper than those reported for the full sample in tables 2
9 We became aware of Smith [2007] while our paper was under review. Smiths [2007] basic resultsand research designare remarkably similar to ours. The recent increase in eventwindow contribution to volatility in our sample is consistent with the evidence in Francis, Schipper, and Vincent [2002], and the Landsman and Maydew [2002] nding that announcement period return volatility increases over time, but it largely postdates their sample periods. 10 Allegations of biased analyst reports surfaced in 2000. In early 2001, New Yorks attorney general, Eliot L. Spitzer, commenced an investigation that culminated in an April 2003 settlement involving $1.4 billion in nes and fraud charges.
994
TABLE 4 Annual Regressions of Calendar-Year Returns on the Four Earnings-Announcement Window Returns, for a Constant Sample of 1,593 Firms over 19962006 Year Intercept Quarter 1 0.649 1.541 0.622 1.084 0.803 0.905 1.573 1.879 0.944 1.194 1.089 1.117 0.36 Quarter 2 1.708 0.926 0.898 2.104 1.072 0.636 2.804 1.713 1.370 1.311 1.141 1.426 0.05 Quarter 3 1.082 1.233 1.501 1.347 1.426 1.638 0.627 0.782 1.458 1.196 1.108 1.218 0.04 Quarter 4 1.390 0.982 0.570 2.267 0.594 0.887 0.588 1.872 1.374 1.089 1.042 1.151 0.38 Abnormal No. of R2 % Obs. 7.4 7.1 3.5 0.2 4.9 3.0 17.2 4.5 12.6 16.9 17.0 8.6 1,593 1,593 1,593 1,593 1,593 1,593 1,593 1,593 1,593 1,593 1,593 1,593
Panel A: Arithmetic returns 1996 0.199 1997 0.294 1998 0.030 1999 0.241 2000 0.105 2001 0.206 2002 0.081 2003 0.568 2004 0.218 2005 0.049 2006 0.167 19962006 mean 0.181 p-value (H 0 = 1) Panel B: Logarithmic returns 1996 0.119 1997 0.195 1998 0.074 1999 0.047 2000 0.034 2001 0.093 2002 0.159 2003 0.381 2004 0.153 2005 0.005 2006 0.120 19962006 mean 0.077 p-value (H 0 = 1)
0.794 1.321 0.719 0.643 0.725 0.861 1.352 0.801 0.826 1.122 0.997 0.924 0.32
1.294 0.868 1.148 1.031 1.108 0.399 1.461 0.825 1.119 1.153 0.886 1.027 0.76
0.823 1.051 1.241 0.721 1.496 1.227 0.790 0.639 1.074 1.255 0.927 1.022 0.79
1.043 0.731 0.571 1.173 0.866 0.694 0.421 0.857 0.958 0.929 0.975 0.838 0.03
9.1 9.1 4.6 3.0 7.7 5.3 9.8 6.2 18.0 20.2 22.0 10.5
1,593 1,593 1,593 1,593 1,593 1,593 1,593 1,593 1,593 1,593 1,593 1,593
This table presents estimates from the regression of calendar-year returns on the four quarterly earnings-announcement returns in the year for a constant sample of rms over the years 1996 to 2006. Calendar-year buy-and-hold returns are computed from daily CRSP returns. Earnings-announcement returns are buy-and-hold returns for the three days surrounding the Compustat announcement date. Firm-years with other than four earnings announcements or with daily return data for fewer than 240 trading days are excluded. Panels A and B present estimates using arithmetic and logarithmic returns, respectively. The p-values are from F -tests that the mean coefcient is 1.0. Abnormal R 2 -square is the regression adjusted R 2 value minus its expectation assuming i.i.d. daily returns. Only rms that have data to estimate the regressions in each of the years 1996 to 2006 are included in the analysis.
and 3, so the increase does not appear to be caused by a change in sample composition. 11
4.8
CONCURRENT MANAGEMENT FORECASTS
The apparent increase in information released in the announcement window appears due only in small part to the post-2001 increase in concurrently
11 The abnormal R 2 values in table 4 generally are larger than for their full-sample counterparts in tables 2 and 3, especially in the later years, possibly due to the 10-year sample selection bias.
995
TABLE 5 Percentage of Firm-Quarters with Management Forecasts Issued in and between Earnings-Announcement Windows In the Earnings-Announcement Window Year 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 Day 0 0.02 0.14 0.26 0.56 1.21 2.24 2.18 8.15 10.53 12.47 14.59 15.46 17.47 Day 1 0.02 0.13 0.19 0.24 0.22 0.17 0.10 0.12 0.17 0.17 0.21 0.15 0.20 Day +1 0.00 0.07 0.04 0.08 0.20 0.32 0.17 0.29 0.25 0.22 0.28 0.24 0.43 Between EarningsAnnouncement Windows 0.56 2.00 3.26 4.09 6.39 6.73 5.19 8.65 7.06 6.46 6.38 5.58 5.37
This table presents the annual frequency of management forecasts as a function of event time. It reports, for each calendar year, the percentage of earnings announcements where a management forecast is issued on the day of the announcement (day 0), on the day before the announcement (day 1), on the day after the announcement (day +1), and outside earnings-announcement windows (i.e., between day +1 of the previous quarter and day 1 of the current quarter). Management forecasts are obtained from the First Call database for all rms with earnings announcements on the CRSPCompustat merged database, and are available from 1994 onwards. All management forecasts in First Call are considered, irrespective of their periodicity (quarterly or annual) and whether the forecast was a number, a range, or a qualitative statement.
released management forecasts, reported by Anilowski, Feng, and Skinner [2007, table 2]. 12 Table 5 reports, for each calendar year, the percentage of earnings announcements where a management forecast is issued on the day of the announcement (day 0), on the day before the announcement (day 1), on the day after the announcement (day +1), and outside the earningsannouncement windows (i.e., between day +1 of the previous quarter and day 1 of the current quarter). Management forecasts are obtained from the First Call database for all rms with earnings announcements on the CRSPCompustat merged database, and are available from 1994 onwards. All management forecasts in First Call are considered, irrespective of their periodicity (quarterly or annual) and whether the forecast is a number, a range, or a qualitative statement. Consistent with the evidence in Anilowski, Feng, and Skinner [2007, table 2], after approximately 2001 a larger fraction of rms release concurrent forecasts. By 2006, more than one in six quarterly earnings announcements are accompanied by a management forecast. 13
12 See also Waymire [1985], Hoskin, Hughes, and Ricks [1986], and Rogers, Skinner, and Van Buskirk [2008]. 13 The effect of the increased First Call coverage of management forecasts, identied by Anilowski, Feng, and Skinner [2007, p. 60], is unclear. This does not explain the observed post-2001 increase in management forecasts between earnings windows, unless First Call has increased its coverage of forecasts only if concurrent with earnings announcements.
996
TABLE 6 Regressions of Calendar-Year Returns on the Four Earnings-Announcement Window Returns, after Excluding Firm-Years in Which a Management Forecast Is Issued Quarter Quarter Quarter Quarter Abnormal No. of Year Intercept 1 2 3 4 R2 % Obs. Panel A: Excludes rm-years with a management forecast in an earnings-announcement window Arithmetic 19942006 mean 0.181 1.246 1.138 1.118 1.154 5.2 4,595 returns p-value (H 0 = 1) 0.03 0.20 0.18 0.39 19942006 median 0.171 1.263 0.961 1.211 1.097 4.5 4,673 p-value (H 0 = 1) 0.02 0.12 0.02 0.02 19941999 mean 0.179 1.126 0.996 1.063 1.241 2.4 5,165 20002006 mean 0.183 1.348 1.260 1.165 1.080 7.5 4,106 Logarithmic 19942006 mean returns p-value (H 0 = 1) 19942006 median p-value (H 0 = 1) 19941999 mean 20002006 mean 0.004 0.050 0.006 0.013 1.149 0.07 1.080 0.00 1.019 1.261 1.072 0.24 1.067 0.59 0.997 1.137 1.139 0.18 1.055 0.59 0.961 1.290 1.074 0.28 1.028 0.12 0.953 1.177 1.157 0.21 1.040 0.12 0.999 1.292 1.129 0.17 1.254 0.02 1.062 1.187 1.161 0.21 1.020 0.12 0.984 1.313 1.051 0.53 1.131 0.59 1.130 0.984 1.143 0.44 1.057 0.02 1.243 1.058 1.042 0.60 1.132 0.12 1.129 0.968 9.6 8.1 6.6 12.3 5.1 4.2 2.3 7.5 9.7 8.6 6.5 12.5 4,595 4,673 5,165 4,106 4,033 4,101 4,481 3,650 4,033 4,101 4,481 3,650
Panel B: Excludes rm-years with management forecasts Arithmetic 19942006 mean 0.197 1.257 returns p-value (H 0 = 1) 0.04 19942006 median 0.183 1.246 p-value (H 0 = 1) 0.02 19941999 mean 0.201 1.153 20002006 mean 0.193 1.346 Logarithmic 19942006 mean returns p-value (H 0 = 1) 19942006 median p-value (H 0 = 1) 19941999 mean 20002006 mean 0.009 0.062 0.022 0.002 1.157 0.05 1.093 0.00 1.025 1.271
This table replicates tables 2 and 3 after excluding rm-years in which management forecasts are issued. Management forecasts are obtained from the First Call database for all rms with earnings announcements on the CRSPCompustat merged database, and are available from 1994 onwards. All management forecasts in First Call are considered, irrespective of their periodicity (quarterly or annual) and whether the forecast was a number, a range, or a qualitative statement. In panel A, only rm-years in which a management forecast is issued in an earnings-announcement window are excluded. In panel B, all rm-years in which a management forecast is issued are excluded, irrespective of whether the forecast is issued within an earnings-announcement window or not.
This result suggests two things. First, the increased frequency of concurrent management forecasts could explain at least part of the sharp increase in the apparent informativeness of earnings announcements after 2000. Second, the abnormal R 2 values reported in tables 2 and 3 could overstate the informativeness of earnings announcements per se. To better understand this issue, panel A of table 6 reports results for a sample that excludes all rm-years in which a management forecast was issued in any of the four earnings-announcement windows. From 1994 (when the First Call data commence) to 2002, there are only minor differences between these results and the full-sample results in tables 2 and 3. For arithmetic returns, the average abnormal R 2 value in the period 1994 to 2006 is 5.2%, marginally lower than the 5.6% for the full sample in table 2. The equivalent estimate for logarithmic returns is marginally higher than for the full sample in table 3 (9.6% vs. 9.3%). Moreover, from untabulated results we nd that the average abnormal R 2 value in the last four sample years, 2003 to 2006, falls, on average, by only 1% when concurrent management forecasts
997
are excluded. These results are not sensitive to the choice of arithmetic versus logarithmic returns, and similar results are obtained from a constantcomposition sample excluding management forecasts during earnings windows. This evidence implies that concurrent management forecasts explain only a small amount of the informativeness of earnings announcements as reported in tables 2 and 3 and of the sharp increase in recent years. To further examine whether the relatively low informativeness of earnings announcements observed in tables 2 and 3 is due to managers prereleasing earnings information through management forecasts, we exclude all rmyears in which a management forecast is issued at any time, either inside or outside the four event windows. The results, reported in panel B of table 6, show that here too the abnormal R 2 values are similar to those for the full sample. On average, over 1994 to 2006, the abnormal R 2 values for arithmetic returns decline from 5.6% in table 2 to 5.2% when rm-years with management forecasts are excluded. A small increase in the average abnormal R 2 value is observed for logarithmic returns, from 9.3% to 9.6%. These results indicate that earnings announcements have relatively low explanatory ability for annual stock return volatility, even when managers do not preempt the announcements with earnings forecasts.
4.9
EFFECT OF SIZE
Earnings announcements are generally viewed as more informative for smaller rms, because they have a sparser information environment. Smaller rms attract lower media coverage and analyst following, resulting in lower information production outside their earnings-announcement windows (Atiase [1985], Collins, Kothari, and Rayburn [1987]). This reasoning addresses only the supply of information to the market, and ignores the likelihood that the demand for information increases in size. Simply put, large rms have more information requiring analysis than do small rms: They tend to have a wider range of products, greater geographical dispersion in operations, more assets, more liabilities, and more customers. If all the activities of a rm are independent, there are no scale economies in information production and we expect rm size to be proportional to the amount of observed information produced, and hence proportional to proxies for information production such as the number of analysts following its stock. We then expect both the unconditional variance of returns and the conditional (on earnings announcements) variance to decline with size at the same rate, leaving the relative informativeness of earnings independent of size. Because the activities of a rm likely are positively correlated (e.g., the reaction to a new product is similar across consumers), following standard asset-pricing logic one might expect information production to be a convex function of size and the relative informativeness of earnings to be a concave function of size. To examine this issue, each year we sort rms into quintiles based on size and then estimate separate calendar-year regressions for each quintile. The results from these size-sorted regressions are presented in panels A through C of table 7 for three size proxies: market capitalization, book value of
998
TABLE 7 Average Abnormal R 2 Values and Slope Coefcients from Annual Regressions of Calendar-Year Returns on the Four Earnings-Announcement Window Returns, for Firms Sorted on Size, Market-to-Book Ratio, Book Leverage, and One-Digit SIC Code Quarter Quarter Quarter Quarter Abnormal No. of Quintile Period Intercept 1 2 3 4 R2 % Obs. Panel A: Market capitalization quintiles Quintile 1 19722006 mean 0.209 1.132 1.158 1.004 1.109 5.6 683 (small) p-value (H 0 = 1) 0.46 0.18 0.96 0.41 19722006 median 0.226 0.897 1.072 1.139 0.961 5.2 644 p-value (H 0 = 1) 0.02 0.10 0.32 0.00 19721989 mean 0.151 0.889 1.176 1.051 0.998 5.5 472 19902006 mean 0.270 1.391 1.140 0.955 1.227 5.7 907 Quintile 2 19722006 mean p-value (H 0 = 1) 19722006 median p-value (H 0 = 1) 19721989 mean 19902006 mean 19722006 mean p-value (H 0 = 1) 19722006 median p-value (H 0 = 1) 19721989 mean 19902006 mean 19722006 mean p-value (H 0 = 1) 19722006 median p-value (H 0 = 1) 19721989 mean 19902006 mean 19722006 mean p-value (H 0 = 1) 19722006 median p-value (H 0 = 1) 19721989 mean 19902006 mean 0.194 0.218 0.174 0.216 0.162 0.169 0.160 0.164 0.140 0.145 0.152 0.127 0.123 0.139 0.129 0.116 1.125 0.13 1.157 0.02 0.922 1.340 1.232 0.01 1.273 0.00 1.177 1.290 1.217 0.00 1.123 0.02 1.235 1.198 1.186 0.01 1.104 0.10 1.259 1.109 1.173 0.38 1.046 0.74 0.806 1.561 1.242 0.00 1.154 0.02 1.233 1.251 1.244 0.03 1.148 0.00 1.245 1.243 1.205 0.07 1.044 0.00 1.067 1.350 1.174 0.00 1.165 0.00 1.155 1.193 1.215 0.05 1.116 0.00 1.036 1.404 1.012 0.89 0.974 0.00 1.038 0.984 1.138 0.19 1.044 0.00 1.185 1.089 1.132 0.22 1.108 0.74 1.031 1.239 1.238 0.01 1.216 0.10 1.121 1.362 1.153 0.10 1.149 0.74 1.158 1.148 1.292 0.00 1.308 0.00 1.298 1.285 1.263 0.00 1.250 0.74 1.265 1.261 1.117 0.24 1.145 0.32 1.110 1.124 1.171 0.08 1.199 0.00 1.273 1.062 1.034 0.76 1.009 0.10 0.945 1.128 1.227 0.03 1.124 0.10 1.231 1.223 1.093 0.32 1.006 0.00 1.091 1.095 1.162 0.09 1.172 0.00 1.137 1.188 1.054 0.58 1.039 0.00 1.125 0.979 1.063 0.52 0.951 0.00 1.186 0.934 0.937 0.60 0.850 0.00 0.784 1.098 1.164 0.27 1.035 0.00 1.001 1.337 1.055 0.60 1.047 0.00 1.052 1.058 6.3 5.2 6.9 5.6 7.6 7.1 7.1 8.1 8.7 9.4 6.9 10.6 7.0 6.1 5.9 8.1 5.5 4.1 5.0 5.9 6.6 6.4 6.2 7.1 7.3 6.4 7.1 7.5 684 644 473 908 684 645 473 908 684 644 473 908 684 644 472 908 650 641 394 921 650 642 394 922 650 642 394 921
(Continued)
Quintile 3
Quintile 4
Quintile 5 (large)
Panel B: Book value of total assets quintiles Quintile 1 19722006 mean 0.215 (small) p-value (H 0 = 1) 19722006 median 0.191 p-value (H 0 = 1) 19721989 mean 0.176 19902006 mean 0.257 Quintile 2 19722006 mean p-value (H 0 = 1) 19722006 median p-value (H 0 = 1) 19721989 mean 19902006 mean 19722006 mean p-value (H 0 = 1) 19722006 median p-value (H 0 = 1) 19721989 mean 19902006 mean 0.169 0.199 0.149 0.190 0.153 0.177 0.150 0.156
Quintile 3

T A B L E 7 Continued Quarter Quarter Quarter Intercept 1 2 3 0.146 1.108 1.270 1.140 0.30 0.02 0.13 0.164 1.114 1.206 1.173 0.00 0.00 0.10 0.145 1.067 1.042 1.063 0.147 1.151 1.511 1.222 0.139 1.049 1.201 1.196 0.68 0.08 0.06 0.159 1.090 1.014 1.220 0.74 0.00 0.32 0.136 0.968 1.155 1.207 0.142 1.135 1.249 1.184 1.186 0.30 0.989 0.02 0.875 1.516 1.167 0.09 1.159 0.00 1.020 1.322 1.282 0.00 1.266 0.00 1.256 1.309 1.114 0.08 1.142 0.00 1.211 1.012 1.155 0.05 1.110 0.74 1.203 1.104 1.070 0.49 1.022 0.74 0.936 1.211 1.370 0.20 1.210 0.00 1.166 1.586 1.092 0.40 1.051 0.10 1.065 1.121 1.236 0.02 1.087 0.0 2 1.162 1.313 1.214 0.00 1.238 0.00 1.115 1.318 1.242 0.01 1.105 0.00 1.212 1.274 1.083 0.39 1.015 0.02 0.967 1.205 1.185 0.11 1.039 0.02 1.121 1.253 1.297 0.07 1.258 0.10 1.235 1.362 1.049 0.56 1.117 0.00 1.054 1.045 1.139 0.08 1.182 0.00 1.174 1.103 1.213 0.05 1.098 0.32 1.213 1.213 1.222 0.02 1.097 0.02 1.168 1.280 1.135 0.09 1.224 0.32 1.102 1.170 1.122 0.14 1.127 0.32 1.118 1.126 1.038 0.68 1.041 0.32 1.063 1.011
999
No. of Obs. 650 642 394 922 650 641 394 921 677 635 453 915 678 636 453 916 678 635 453 916 678 636 453 916 677 635 453 915 663 628 444 894 663 628 445 895
(Continued)
Quintile Quintile 4
Quintile 5 (large)
Period 19722006 mean p-value (H 0 = 1) 19722006 median p-value (H 0 = 1) 19721989 mean 19902006 mean 19722006 mean p-value (H 0 = 1) 19722006 median p-value (H 0 = 1) 19721989 mean 19902006 mean
Quarter 4 1.075 0.46 1.131 0.00 1.174 0.971 1.061 0.61 1.133 0.00 1.124 0.994 0.956 0.69 0.787 0.00 0.817 1.104 1.270 0.01 1.204 0.00 1.289 1.250 1.106 0.32 1.045 0.00 1.059 1.157 1.124 0.25 1.066 0.00 1.189 1.055 1.058 0.55 1.078 0.00 1.133 0.979 1.070 0.57 1.037 0.00 1.003 1.141 1.036 0.71 1.037 0.00 1.040 1.033
Abnormal R2 % 6.1 5.8 3.5 8.9 6.1 6.3 5.9 6.2 5.2 5.2 5.1 5.4 7.8 8.2 8.2 7.4 7.4 6.9 7.1 7.6 8.0 7.6 7.0 9.0 6.8 6.1 5.8 7.8 5.9 6.1 6.2 5.6 8.0 6.9 7.6 8.3
Panel C: Book value of equity quintiles Quintile 1 19722006 mean 0.216 (small) p-value (H 0 = 1) 19722006 median 0.191 p-value (H 0 = 1) 19721989 mean 0.172 19902006 mean 0.263 Quintile 2 19722006 mean 0.176 p-value (H 0 = 1) 19722006 median 0.185 p-value (H 0 = 1) 19721989 mean 0.157 19902006 mean 0.196 Quintile 3 19722006 mean 0.157 p-value (H 0 = 1) 19722006 median 0.188 p-value (H 0 = 1) 19721989 mean 0.149 19902006 mean 0.167 Quintile 4 19722006 mean 0.145 p-value (H 0 = 1) 19722006 median 0.159 p-value (H 0 = 1) 19721989 mean 0.151 19902006 mean 0.139 Quintile 5 19722006 mean 0.133 (large) p-value (H 0 = 1) 19722006 median 0.144 p-value (H 0 = 1) 19721989 mean 0.137 19902006 mean 0.129 Panel D: Market-to-book ratio quintiles Quintile 1 19722006 mean 0.234 (small) p-value (H 0 = 1) 19722006 median 0.216 p-value (H 0 = 1) 19721989 mean 0.203 19902006 mean 0.267 Quintile 2 19722006 mean 0.176 p-value (H 0 = 1) 19722006 median 0.185 p-value (H 0 = 1) 19721989 mean 0.167 19902006 mean 0.186
1000

T A B L E 7 Continued Quarter Quarter Quarter Intercept 1 2 3 0.163 1.169 1.153 1.270 0.00 0.10 0.01 0.185 1.182 1.111 1.206 0.00 0.00 0.10 0.157 1.080 1.140 1.299 0.169 1.263 1.166 1.239 0.133 1.152 1.084 1.066 0.05 0.19 0.50 0.139 1.159 0.988 1.071 0.74 0.32 0.74 0.132 1.042 1.084 0.906 0.135 1.269 1.085 1.236 0.125 1.090 1.090 1.120 0.32 0.34 0.27 0.115 1.061 1.099 0.954 0.02 0.74 0.00 0.114 0.935 1.067 1.057 0.137 1.254 1.114 1.187 0.176 0.181 0.142 0.213 0.177 0.158 0.156 0.200 0.162 0.186 0.158 0.167 0.154 0.175 0.162 0.146 0.157 0.190 0.149 0.166 1.414 0.06 1.080 0.32 1.131 1.715 1.181 0.06 1.158 0.10 1.022 1.350 1.196 0.01 1.220 0.10 1.206 1.186 1.110 0.24 0.988 0.00 1.088 1.133 0.985 0.86 0.995 0.02 0.889 1.085 1.248 0.08 1.016 0.02 1.159 1.342 1.244 0.01 1.114 0.00 1.210 1.280 1.084 0.28 1.084 0.00 1.017 1.154 1.232 0.02 1.088 0.00 1.160 1.308 1.232 0.09 1.173 0.74 1.072 1.401 1.083 0.36 1.183 0.74 1.171 0.989 1.157 0.05 1.153 0.32 1.124 1.192 1.276 0.01 1.342 0.02 1.209 1.347 1.200 0.07 1.274 0.10 1.198 1.202 1.081 0.40 1.017 0.02 1.002 1.164
Quintile Quintile 3
Quintile 4
Quintile 5 (large)
Period 19722006 mean p-value (H 0 = 1) 19722006 median p-value (H 0 = 1) 19721989 mean 19902006 mean 19722006 mean p-value (H 0 = 1) 19722006 median p-value (H 0 = 1) 19721989 mean 19902006 mean 19722006 mean p-value (H 0 = 1) 19722006 median p-value (H 0 = 1) 19721989 mean 19902006 mean
Quarter 4 1.239 0.04 1.377 0.00 1.146 1.337 1.122 0.31 1.118 0.00 1.026 1.224 1.037 0.75 0.885 0.00 1.081 0.991 1.116 0.39 1.019 0.00 1.021 1.215 1.129 0.27 0.974 0.00 1.112 1.147 1.107 0.19 1.079 0.00 1.158 1.052 1.095 0.29 1.040 0.00 0.973 1.223 1.029 0.83 0.916 0.00 1.079 0.975
Abnormal R2 % 7.3 7.5 7.1 7.6 6.0 5.3 4.9 7.2 5.0 4.8 4.3 5.7 7.7 7.6 7.8 7.5 6.7 5.6 6.9 6.6 7.2 7.9 6.7 7.7 7.8 6.9 7.3 8.2 4.8 4.1 4.4 5.3
No. of Obs. 663 628 445 894 663 628 445 895 663 628 444 894 700 622 445 970 638 623 446 841 669 623 445 906 669 623 446 906 669 623 445 905
Panel E: Debt-to-equity ratio Quintile 1 19722006 mean (small) p-value (H 0 = 1) 19722006 median p-value (H 0 = 1) 19721989 mean 19902006 mean Quintile 2 19722006 mean p-value (H 0 = 1) 19722006 median p-value (H 0 = 1) 19721989 mean 19902006 mean Quintile 3 19722006 mean p-value (H 0 = 1) 19722006 median p-value (H 0 = 1) 19721989 mean 19902006 mean Quintile 4 19722006 mean p-value (H 0 = 1) 19722006 median p-value (H 0 = 1) 19721989 mean 19902006 mean Quintile 5 19722006 mean (large) p-value (H 0 = 1) 19722006 median p-value (H 0 = 1) 19721989 mean 19902006 mean
(Continued)
assets, and book value of equity. Returns are arithmetic (similar results apply to logarithmic returns). Extreme observations are not deleted. For each of the three size proxies, the abnormal adjusted R 2 value averaged across all years is best described as a concave function of size. In each case, the

T A B L E 7 Continued Panel F: Industry classication (one-digit SIC code) Quarter Quarter Quarter Period Intercept 1 2 3 Agriculture, forestry, shing, mining, and construction (0,1) 19722006 mean 0.182 1.557 1.029 0.903 p-value (H 0 = 1) 0.28 0.89 0.55 19722006 median 0.159 0.867 0.787 0.900 p-value (H 0 = 1) 0.02 0.74 0.02 19721989 mean 0.178 0.990 0.871 0.882 19902006 mean 0.185 2.157 1.196 0.925 Manufacturing (2,3) 19722006 mean p-value (H 0 = 1) 19722006 median p-value (H 0 = 1) 19721989 mean 19902006 mean 0.166 0.169 0.150 0.183 1.165 0.05 1.146 0.00 1.066 1.270 1.130 0.03 1.100 0.00 1.184 1.072 1.188 0.00 1.255 0.02 1.154 1.225 1.197 0.25 0.975 0.02 1.028 1.375 1.120 0.28 0.968 0.02 1.105 1.137 1.149 0.20 1.109 0.32 1.418 0.865 1.308 0.02 1.191 0.74 1.483 1.122
1001
Quarter 4 1.317 0.08 1.176 0.00 1.475 1.150 1.119 0.24 1.036 0.00 1.034 1.208 0.859 0.33 0.871 0.00 0.891 0.826 1.001 1.00 0.859 0.00 0.998 1.003 1.205 0.13 1.045 0.00 1.198 1.214 1.048 0.72 1.020 0.00 1.052 1.045
Abnormal R2 % 5.8 2.7 4.8 6.9 7.0 6.0 7.1 6.8 5.6 5.3 5.5 5.8 8.1 7.7 6.5 9.7 6.5 6.1 7.0 5.9 7.1 6.9 7.5 6.7
No. of Obs. 241 255 209 276 1549 1486 1250 1865 308 319 233 387 355 345 242 474 559 446 247 889 453 385 182 740
Transportation, communications, and utilities (4) 19722006 mean 0.154 1.057 1.181 p-value (H 0 = 1) 0.59 0.24 19722006 median 0.173 1.060 1.099 p-value (H 0 = 1) 0.02 0.74 19721989 mean 0.147 1.075 0.851 19902006 mean 0.160 1.039 1.530 Wholesale and retail trade (5) 19722006 mean 0.154 p-value (H 0 = 1) 19722006 median 0.141 p-value (H 0 = 1) 19721989 mean 0.163 19902006 mean 0.144 1.070 0.54 1.051 0.00 1.063 1.077 1.263 0.06 1.260 0.74 0.999 1.543 1.146 0.16 1.195 0.02 0.988 1.312 1.169 0.22 1.069 0.02 1.141 1.199
Finance, insurance, and real estate (6) 19722006 mean 0.150 0.885 p-value (H 0 = 1) 0.44 19722006 median 0.162 0.872 p-value (H 0 = 1) 0.00 19721989 mean 0.125 0.705 19902006 mean 0.176 1.075 Services (7,8) 19722006 mean p-value (H 0 = 1) 19722006 median p-value (H 0 = 1) 19721989 mean 19902006 mean 0.190 0.189 0.177 0.204 1.158 0.23 1.082 0.32 0.969 1.357
This table presents estimates from annual regressions of calendar-year returns on the four earningsannouncement window returns, for rms sorted on size, market-to-book ratio, book leverage, and one-digit SIC code. Regressions are as described in tables 2 and 3, but are estimated separately for stocks sorted annually on variables that potentially proxy for the rms information environment. Panels A, B, and C present results for size quintiles based on prior year-end market capitalization, book value of total assets, and book value of equity, respectively. Panel D presents results for quintiles based on the prior year-end market-to-book ratio, panel E presents results for quintiles based on the prior year-end debt-to-equity ratio (dened as the book value of total long-term debt divided by the book value of equity), and panel F presents results for rms classied by one-digit SIC code. The table reports the means and medians of estimates across the yearly regressions. The p-values and abnormal R 2 are as dened in tables 2 and 3. Returns are arithmetic.
smallest-size quintile exhibits the lowest abnormal R 2 value: 5.6%, 5.5%, and 5.2%. The largest-size quintile exhibits the second-lowest abnormal R 2 value for two proxies and the third-lowest for the other. Similar concavity is observed in the size quintiles medians, suggesting it is not due to
1002
outliers. Thus, while small stocks exhibit higher return volatility at earnings announcements, their announcements appear to contribute a lower proportion of their annual return volatility. These results could be affected by three-day returns for small rms being noisier, due to market pricing issues such as bidask bounce. However, in the following subsection we report a similar concavity for portfolios formed on market-to-book ratios.
4.10
EFFECT OF THE MARKET-TO-BOOK RATIO, LEVERAGE, AND INDUSTRY
The informativeness of earnings is likely to depend on a variety of rm characteristics. We study market-to-book (MTB) ratios, leverage, and industry, using arithmetic returns. The likely effect of MTB is unclear. To the extent that MTB reects unbooked growth options, earnings are likely to be less informative about changes in rm value, which predicts that the abnormal R 2 decreases with MTB. To the extent that MTB reects the unbooked intellectual property of mature, nongrowth businesses with long-term earnings streams, such as Coca Cola and Microsoft, earnings are likely to be more informative about changes in rm value, which predicts that the abnormal R 2 increases with MTB. Panel D of table 7 reports results for MTB quintiles. The relation between MTB and abnormal R 2 appears concave, consistent with both of the above effects existing in the data. Panel E of table 7 reports results for quintiles formed on book leverage, dened as the book value of long-term debt divided by the book value of equity. There is no evident relation between book leverage and abnormal R 2 , possibly because leverage has a similar effect on the volatilities of stock returns and earnings. The largest leverage quintile exhibits the lowest abnormal R 2 value, 4.8%. To provide better insight into the effects of rm characteristics on earnings informativeness, we also study industry effects. It seems plausible that informativeness differs between (say) manufacturing and nance, due to a greater proportion of transitory earnings components in the latter. Panel F of table 7 reports results for seven broad industry groupings based on onedigit Standard Industrial Classication (SIC) codes. Wholesale and retail rms exhibit the highest abnormal R 2 values, and transportation, communication, and utilities exhibit the lowest. The surprising result is the consistency of the relative informativeness of earnings across industries: The range of abnormal R 2 values is only 5.6% to 8.1%.
4.11
TIME-SERIES REGRESSIONS
As a specication check, we also estimate rm-level time-series regressions. When the minimum required number of years with return data is set to 20, the sample size is 1,427 rms and the mean (median) rms estimated abnormal R 2 is 8.9% (6.5%) for arithmetic returns and 9.7% (7.6%) for logarithmic returns. These results could be affected by survivor bias, though similar results are obtained for larger samples when the data requirement is reduced to 15 or 10 years. Overall, the results are not substantially different from those of the cross-sectional regressions.
1003
4.12
DIFFERENT EVENT WINDOWS
The results are not sensitive to the length of the chosen event window. For example, with a two-day window (day 0 and day +1, where day 0 is the Compustat earnings-announcement date), the average abnormal R 2 over the full period 1972 to 2006 is 4.5% for arithmetic returns (compared to 5.9% in table 2 for a three-day window) and 7.2% for logarithmic returns (compared to 9.3% in table 3). The reduction occurs because, as reported in the following section, there is a negative abnormal R 2 value on day 1, which is dropped from this calculation. 14 Lower abnormal R 2 values also are observed when the event window is extended to 5 days (2 to +2).
5. Information in Adjacent Nonevent Windows

We next examine information ows in the periods surrounding the earnings-announcement windows. We study various event-time windows that are distanced in event time from the earnings-announcement windows by plus or minus 0, 5, 10, 15, 20, 25, or 30 trading days. For example, when the distance from the earnings-announcement window is +10 trading days, the calendar-year return is regressed annually on returns in the four nonearnings-announcement windows comprising days +9 to +11 relative to the four annual Compustat earnings-announcement dates. The period studied is approximately the six weeks before, the week of, and the six weeks after the quarterly earnings announcementa period of approximately one quarter in length, centered on the announcement day. As previously, only rms with exactly four earnings announcements in a calendar year are included, and we require the three-day event windows to lie entirely within the calendar year. The regressions are estimated annually from arithmetic returns, and the average of the abnormal adjusted R 2 values and slope coefcients are plotted in gure 2. Not surprisingly, the abnormal adjusted R 2 in panel A of gure 2 is highest when the three-day event window corresponds to the earningsannouncement window. It is slightly negative in each of the six weeks before and six weeks after the announcement. 15 It is lowest immediately prior to the earnings-announcement window, that is, during the three-day window commencing ve trading days before the earnings-announcement window.
14 It is well known that many earnings announcements occur after the close of trading on the Compustat announcement date, so the window must include both days 0 and 1 to avoid biasing the estimated abnormal R 2 downward. Berkman and Truong [2008] report that the frequency of after-hours earnings announcements exceeds 40% in recent years. 15 If returns are i.i.d. and hence the regression slopes are unity, the positive abnormal R 2 value in the event week would mechanically induce a small negative abnormal R 2 value in the average nonevent week. The comparatively small abnormal R 2 values in the adjacent nonevent windows explain why the event-window abnormal R 2 is relatively insensitive to the event window widthas the window widens, the contribution of the window to annual return volatility rises at approximately the same rate as its i.i.d. benchmark. Our results are similar to those in Beaver [1968] and Morse [1981], who employ different but related research designs and study earlier and smaller samples.
1004
FIG. 2.Average slope coefcients and adjusted abnormal R 2 values from annual regressions of calendar-year arithmetic returns on arithmetic returns in four three-day windows that vary in event time relative to the four quarterly earnings-announcement windows. The event-time windows are distanced in event time from the earnings-announcement windows by plus or minus 0, 5, 10, 15, 20, 25, or 30 trading days. For example, when the distance from the earningsannouncement window is +10 trading days, the calendar-year return is regressed annually on returns in the four non-earnings-announcement windows comprising days +9 to +11 relative to the four annual Compustat earnings-announcement dates. When the distance from the announcement window is 0 trading days, calendar-year return is regressed annually on returns in the four earnings-announcement windows themselves. Statistics are averages of regression parameters across years. Calendar-year arithmetic buy-and-hold returns are computed from daily CRSP returns. Earnings-announcement window returns and non-announcement window returns are arithmetic buy-and-hold returns over the three days. The sample is all rm-years with available data on the quarterly Compustat and daily CRSP les. Firm-years with other than four earnings announcements or with daily returns data for fewer than 240 trading days are excluded. The abnormal adjusted R 2 is the regression adjusted R 2 minus its expectation assuming i.i.d. daily returns.
1005
This result holds in both the 1972 to 1989 and the 1990 to 2005 periods, and for logarithmic as well as arithmetic returns. The result is consistent with private information production being comparatively depressed immediately around earnings announcements, for example, because public announcements decrease investors incentives to produce new information (i.e., crowd it out), because rms impose restrictions on insider trading or on earnings guidance updates around earnings announcements, or simply because information production opportunities are reduced in that period. 16 Panel B of gure 2 reports the average slope coefcients from three-day event windows around the earnings-announcement dates. The coefcients for all the non-earnings-announcement windows are substantially less than unity, implying systematic reversals, consistent with initial market overreaction. The average quarters slope coefcient is lowest for the three-day event window in the week immediately before the earnings-announcement window, which seems to be the period when returns are more driven by liquidity and other transitory shocks than by price-relevant information. 17
6. Analyst Earnings Forecasts and Management Forecasts

Our regression design can be adapted to quantify the relative informativeness of some other variables, such as dividends, 10-Q lings, and earnings forecasts. This section reports some results on earnings forecasts by analysts and managers. The evidence that information arrival appears slightly lower than normal both before and after earnings announcements seems somewhat inconsistent with the view that security analysts produce substantial new information. In the weeks prior to earnings announcements, analysts might be expected to release above-normal amounts of information. In the weeks following earnings announcements, analysts commonly revise their forecasts of future earnings. Our results suggest that much forecast revision immediately after earnings announcements consists of incorporating the new information in the announcement into the forecasts, rather than producing new information.
6.1
EARLIEST POSTANNOUNCEMENT ANALYST FORECAST REVISION
We initially investigate abnormal price volatility at the earliest analyst forecast revision after each of the four quarterly earnings announcements. The
See, for example, Kim and Verrecchia [1997]. The higher abnormal R 2 values and slope coefcients in regressions based on earningsannouncement windows relative to those based on non-earnings-announcement windows could potentially reect differences in return-measurement errors across these event windows. If returns in non-earnings-announcement windows are more affected by stale prices, bidask bounce, or other noise then the slope coefcients and adjusted R 2 values would be more attenuated in these periods.
17
16
1006
objective is to better understand the result that information arrival appears to be slightly lower than normal after earnings announcements, a period when analysts might be expected to revise their forecasts of future earnings in light of the earnings outcome and other information released in the announcement window. Analyst forecast data are available from the I/B/E/S detailed unadjusted data set from 1985 onwards, and 2006 is the last year with sufcient data. Forecast event-window returns are arithmetic buy-and-hold returns for the three days surrounding the rst analyst forecast revision after the previous quarters earnings announcement. There must be at least one forecast revision in each calendar quarter for a rm/year to be included in the sample. Forecasts issued within two trading days after the earnings announcement are excluded, to avoid overlap between the three-day earnings and threeday forecast windows. Results are sensitive to a small number of extreme outliers, so we report results for a sample that excludes, in each calendar year, the extreme 1% on either side of the calendar-year returns, earningsannouncement returns, and forecast event-window returns. Deletion of an individual quarter requires deletion of the entire rm/year from the regression. The resulting sample size peaks at 1,112 in 1992, falls substantially over time, and averages 477 rms per year. Table 8, panel A reports that the average abnormal adjusted R 2 value from annual regressions of calendar-year returns on the four forecast revision window returns is 0.4%. The earliest forecast revision after earnings announcements therefore is not a period of unusual price volatility, consistent with the evidence in gure 2. One possibility is that the rst postannouncement forecast revision typically incorporates the old information released in earnings, not new information, though we note that (particularly for small stocks) many of the rst postannouncement revisions occur considerably later. Each of the four regression slopes is less than unity, though none is signicantly less. Panel B presents results for the earnings-announcement windows of the same sample of rms/years as in panel A. The average abnormal adjusted R 2 for this sample is lower than for the full sample (4.5% vs. 5.8% in table 2 for 1985 to 2006), suggesting that rms receiving forecast revisions each quarter are those whose earnings are slightly less informative than usual. To include forecasts issued on the rst trading day after the earnings announcement, we reduce the forecast event window to its announcement day. Forecasts on the same day as earnings announcements are excluded. In all other respects the analysis is as reported in table 8. The untabulated results are similar to those for three-day windows. The sample size increases slightly, from an average of 477 rms per year to 556. The average abnormal adjusted R 2 from annual regressions of calendar-year returns on forecast revision returns is 0.0%. Caution is advised in interpreting results from single-day windows, because the events could have occurred after the close of trading.
1007
TABLE 8 Annual Regressions of Arithmetic Calendar-Year Returns on Arithmetic Returns in the Window Surrounding the Earliest Earnings Analyst-Forecast Revision Each Quarter and on Earnings-Announcement Window Returns for the Same Sample Period Intercept Quarter Quarter Quarter Quarter Abnormal 1 2 3 4 R2 % No. of Obs.
Panel A: Calendar-year returns regressed on the earliest analyst quarterly forecast revision window returns 19852006 mean 0.160 0.802 0.913 0.803 0.989 0.4 477 p-value (H 0 = 1) 0.12 0.61 0.23 0.96 19852006 median 0.173 0.863 0.850 0.706 1.090 0.8 369 p-value (H 0 = 1) 0.40 0.10 0.10 0.40 19851995 mean 0.151 0.639 0.952 0.814 0.847 1.3 686 19962006 mean 0.170 0.965 0.874 0.793 1.132 0.6 269 Panel B: Calendar-year returns regressed on earnings-announcement window returns, analyst earnings forecast sample 19852006 mean 0.144 1.069 1.078 0.926 1.051 4.5 p-value (H 0 = 1) 0.56 0.55 0.52 0.67 19852006 median 0.151 0.987 1.092 1.041 0.958 5.2 p-value (H 0 = 1) 1.00 0.10 0.40 0.40 19851995 mean 0.144 1.006 1.049 1.052 1.008 4.7 19962006 mean 0.145 1.132 1.108 0.800 1.094 4.3
477 369 686 269
This table presents estimates from annual regressions of arithmetic calendar-year returns on arithmetic returns in the window surrounding the earliest earnings analyst-forecast revision each quarter and on earnings-announcement window returns for the same sample. Event-window returns are arithmetic buy-and-hold returns for the three days surrounding the rst analyst earnings forecast revision after the previous quarters earnings announcement, and surrounding the Compustat earnings announcement date. Analyst forecast data are from the I/B/E/S detailed unadjusted data set. Overlapping forecast and earnings-announcement windows are excluded, as are rm-years with other than four earnings announcements, four forecast event windows, or daily returns data in fewer than 240 trading days. The sample excludes, in each calendar year, the extreme 1% on either side of the calendar-year returns, earnings-announcement window returns, and forecast window returns. Deletion of an individual quarter requires deletion of the entire rm/year. Panel A presents results for forecast event-window returns, while panel B presents results for earnings-announcement returns with the same sample. The table presents means across years, FamaMacBeth t-statistics, and p-values from F -tests that the mean coefcient is 1.0 and sign tests that the frequency of coefcients exceeding 1.0 is 0.5. Abnormal R 2 is the regression adjusted R 2 value minus its expectation assuming i.i.d. daily returns.
6.2
LATEST PREANNOUNCEMENT ANALYST FORECAST REVISION
Forecast revisions made before earnings announcements appear to be associated with more price revision than those made after. Table 9 investigates abnormal price volatility at the latest analyst forecast revision to take before each of the four quarterly earnings announcements. The procedure followed is the same as for postannouncement forecasts as reported in table 8. Here too the sample size peaks in 1992. There are insufcient observations in 2000 to 2002 to estimate regressions. In contrast with the postannouncement forecasts, in panel A the four latest forecast revisions before earnings announcements exhibit an average abnormal adjusted R 2 of 4.7%. This increases substantially in recent years, indicating that reduced analyst activity is not an explanation for the increase in relative informativeness of earnings reported above. Panel B reports a comparatively large average earnings-announcement day abnormal R 2 for this sample of 9.3%.
1008
TABLE 9 Annual Regressions of Arithmetic Calendar-Year Returns on Arithmetic Returns in the Window Surrounding the Latest Analyst Earnings-Forecast Revision Each Quarter and on Earnings-Announcement Window Returns for the Same Sample Period Intercept Quarter 1 Quarter 2 Quarter 3 Quarter 4 Abnormal R2 % No. of Obs.
Panel A: Calendar-year returns regressed on the latest analyst forecast revision window returns 19852006 mean 0.168 1.288 1.629 1.226 1.124 4.7 p-value (H 0 = 1) 0.31 0.14 0.32 0.61 19852006 median 0.176 1.073 1.490 1.287 1.009 4.4 p-value (H 0 = 1) 1.00 0.08 1.00 0.37 19851995 mean 0.139 1.232 1.159 1.200 1.103 3.0 19962006 mean 0.207 1.363 2.275 1.262 1.154 7.0 Panel B: Calendar-year returns regressed on earnings-announcement window returns, analyst earnings forecast sample 19852006 mean 0.170 1.198 1.471 1.403 1.263 9.3 p-value (H 0 = 1) 0.07 0.02 0.07 0.20 19852006 median 0.158 1.177 1.298 1.368 1.091 8.3 p-value (H 0 = 1) 0.37 0.01 0.37 0.37 19851995 mean 0.149 1.108 1.446 1.601 1.305 10.5 19962006 mean 0.198 1.321 1.506 1.132 1.205 7.6
513 381 671 296
513 381 671 296
This table presents estimates from annual regressions of arithmetic calendar-year returns on arithmetic returns in the window surrounding the latest analyst earnings-forecast revision each quarter and on earnings-announcement window returns for the same sample. Forecast event-window returns are arithmetic buy-and-hold returns for the three days surrounding the latest analyst earnings forecast revision after the previous quarters earnings announcement. Earnings-announcement returns are similarly estimated for the three-day window around the Compustat earnings announcement date. Analyst forecast data are from the I/B/E/S detailed unadjusted data set. Forecasts whose window overlaps the earnings window are excluded, as are rm-years with other than four earnings announcements, four forecast event windows, or daily returns data for fewer than 240 trading days. There are fewer than 12 observations in 2000, 2001, and 2002, so these years are excluded. Panel A presents results for forecast event-window returns, while panel B presents results for earnings-announcement returns with the same sample. The table presents means across years, FamaMacBeth t-statistics, and p-values from F -tests that the mean coefcient is 1.0 and sign tests that the frequency of coefcients exceeding 1.0 is 0.5. Abnormal R 2 is the regression adjusted R 2 value minus its expectation assuming i.i.d. daily returns.
6.3
MANAGEMENT FORECASTS
The results reported in table 6 indicate that abnormal R 2 values for earnings-announcement windows decline when concurrent management forecasts are excluded, implying that management forecasts are likely to be informative. To examine the relative informativeness of management forecasts, we modify regression (1) to a quarterly setting, as there are very few rm-years with management forecasts in every calendar quarter. We regress calendar-quarter returns on returns in the three-day windows surrounding management forecast dates obtained from the First Call database. To avoid confounding effects, we consider only management forecasts released outside the three-day earnings-announcement windows and only the last forecast in the rm-quarter, if multiple forecasts are issued. A potential limitation of the quarterly regressions is that any initial underor overreaction is assumed to be corrected by the end of that calendar quarter.

TABLE 10 Regressions of Quarterly Returns on Returns in the Management Forecast Window Arithmetic Returns Logarithmic Returns
1009
Calendar Year Mgmt Mgmt and Forecast Abnormal No. of Forecast Abnormal No. of Obs. Intercept Return R 2 (%) Obs. Quarter Intercept Return R 2 (%) 19981 0.114 1.177 25.6 449 0.088 1.204 35.3 449 19982 0.076 1.130 32.7 493 0.107 1.260 34.1 493 19983 0.235 0.880 34.8 546 0.310 1.221 38.4 546 19984 0.169 0.973 16.7 738 0.121 0.920 22.1 738 19991 0.045 1.061 8.0 720 0.102 1.166 27.8 720 19992 0.141 1.096 22.3 596 0.100 0.944 27.4 596 19993 0.077 1.122 23.5 576 0.123 1.235 27.3 576 19994 0.196 1.555 11.6 609 0.084 1.126 22.7 609 20001 0.106 1.232 14.3 571 0.036 1.060 21.0 571 20002 0.083 1.002 22.4 454 0.140 1.151 24.8 454 20003 0.020 1.205 26.3 543 0.019 1.411 42.3 543 20004 0.096 0.850 15.5 900 0.163 1.432 27.6 900 20011 0.001 0.861 8.3 1,210 0.051 1.187 19.5 1,210 20012 0.173 0.964 7.5 993 0.118 0.837 11.5 993 20013 0.187 0.948 16.8 956 0.250 1.485 23.9 956 20014 0.287 1.609 12.3 1,143 0.208 1.059 11.7 1,143 20021 0.053 1.191 35.3 1,046 0.029 1.254 32.6 1,046 20022 0.071 1.362 24.9 969 0.118 1.511 30.6 969 20023 0.202 0.815 23.1 1,004 0.259 1.150 28.7 1,004 20024 0.103 0.939 8.6 1,230 0.071 0.809 17.7 1,230 20031 0.026 0.983 14.6 989 0.050 1.075 19.2 989 20032 0.261 1.738 20.0 1,028 0.211 1.187 19.6 1,028 20033 0.112 1.218 18.9 909 0.090 1.028 24.0 909 20034 0.147 1.288 22.5 1,162 0.125 1.088 25.3 1,162 20041 0.055 1.151 23.9 1,152 0.041 1.106 24.8 1,152 20042 0.017 1.181 29.6 928 0.004 1.150 31.7 928 20043 0.048 0.894 20.6 1,033 0.063 1.014 24.5 1,033 20044 0.146 1.111 18.0 1,263 0.126 0.904 20.4 1,263 20051 0.042 1.112 25.0 1,010 0.057 1.193 25.9 1,010 20052 0.029 1.131 24.9 978 0.017 1.118 24.0 978 20053 0.036 1.040 20.9 895 0.024 1.063 24.5 895 20054 0.028 1.031 21.4 1,260 0.018 1.039 26.5 1,260 20061 0.118 1.224 20.6 1,043 0.101 1.065 22.4 1,043 20062 0.055 1.040 21.7 978 0.069 1.133 24.9 978 20063 0.022 0.849 23.2 977 0.013 0.857 26.2 977 20064 0.089 1.044 18.9 1,266 0.078 0.989 21.2 1,266 Mean 0.033 1.111 20.4 906 0.005 1.123 25.3 906 0.00 0.00 p-value (H 0 = 1) Median 0.029 1.103 21.2 978 0.017 1.122 24.8 978 0.00 0.00 p-value (H 0 = 1)
Calendar-quarter returns from 1998 to 2006 are regressed each quarter on three-day returns around the last management forecast in First Call issued in the quarter. Forecasts issued within a three-day earnings-announcement window are excluded. Calendar-quarter arithmetic buy-and-hold or logarithmic returns are computed from daily CRSP returns. Firm-quarters with daily returns data available for fewer than 59 trading days are excluded. Management forecast returns are arithmetic buy-and-hold (or logarithmic) returns for the three-day forecast window. The p-values are from F -tests that the mean coefcient is 1.0 and sign tests that the frequency of coefcients exceeding 1.0 is 0.5. Abnormal adjusted R 2 is the regression adjusted R 2 value minus its expectation assuming i.i.d. daily returns.
Table 10 reports average abnormal R 2 values of 20% to 25% for arithmetic and logarithmic returns. For comparability, regressions of calendarquarter returns on earnings-announcement returns in that quarter show average abnormal R 2 values of 3.5% for arithmetic returns and 4.5% for
1010

TABLE 11 Summary Statistics for Abnormal Volume at Earnings Announcements 10th 25th 50th 75th 90th Mean Percentile Percentile Percentile Percentile Percentile Year (%) (%) (%) (%) (%) (%) 19722006 mean 1.80 1.43 0.33 1.11 3.07 5.63 19721989 mean 1.45 1.45 0.48 0.81 2.58 4.95 19902006 mean 2.18 1.40 0.16 1.43 3.58 6.35 1.48 1.49 1.47 1.30 1.39 1.20 0.44 0.57 0.31 0.03 0.32 0.27 0.89 0.63 1.17 1.56 1.11 2.04 2.72 2.27 3.20 3.65 3.04 4.31 5.15 4.52 5.82 6.39 5.52 7.30
All rms
December scal 19722006 mean 1.55 year-end rms 19721989 mean 1.24 19902006 mean 1.88 Non-December scal yearend rms 19722006 mean 2.26 19721989 mean 1.76 19902006 mean 2.79
Abnormal volume is total volume in the four earnings-announcement windows as a percentage of annual volume, minus its expectation assuming daily volume is i.i.d. Announcement windows are days 1 to +1, where day 0 is the Compustat announcement date. Firm-years with other than four earnings announcements or with daily volume data for fewer than 240 trading days are excluded. The table presents time-series averages of annual summary statistics for abnormal volume. The results are presented for the full sample and separately for rms with December and non-December scal year-ends.
logarithmic returnsresults that are comparable to those obtained above from regressions of annual returns on the four event-window returns. The comparison of the informativeness of management forecasts with earnings announcements is striking and informative. Earnings announcements are required quarterly, and are based largely on backward-looking events. In contrast, managers presumably self-select to issue forecasts when they believe they contain substantial forward-looking information. Consequently, management forecasts are less frequent but, when made, they are considerably more informative than earnings announcements.
7. Trading Volume Results

We also examine the abnormal trading volume during the earningsannouncement windows. For each rm-year, we rst compute the trading volume at each of the four quarterly earnings announcements as a percentage of total annual trading volume. We then subtract an estimated 4.8% normal volume, assuming that daily trading volume is i.i.d. across time. We take no position on the interpretation of abnormal volume. 18 Table 11 reports summary statistics for this abnormal trading volume measure. The mean across all years of abnormal volume during the four threeday event windows taken together is 1.80% of annual volume, or only 0.45% per individual quarterly announcement window. The median abnormal volume during the four three-day event windows taken together is even lower, at
18 See, for example, Beaver [1968], Karpoff [1987], Holthausen and Verrecchia [1990], Kim and Verrecchia [1991a, b, 1994], Lee [1992], Harris and Raviv [1993], Kandel and Pearson [1995], Bamber, Barron, and Stober [1999], Odean [1999], Baker and Stein [2004], and Dey and Radhakrishna [2007].
1011
1.11% (0.28% per quarter), consistent with right skew in volume (Bamber, Christensen, and Gaver [2000]). The small average percentage increase in volume at earnings announcements could indicate that earning announcements are not a major source of new information to investors, but it also is consistent with other explanationsfor example, that most trading during the year occurs for reasons unrelated to new information, such as liquidity, taxes, and portfolio rebalancing. The mean abnormal volume increases between subperiods from 1.45% of annual volume to 2.18%; for the average quarterly announcement, this is an increase from 0.36% to 0.55% of annual volume. In untabulated results, the 16 post-1990 years exhibit the 13 highest mean proportions of annual volume. More specically, consistent with the results in tables 2 and 3, the years with the highest abnormal volume are the last three sample years, 2004 to 2006. In these years, the average abnormal trading volumes are 2.69%, 2.95%, and 3.10%, respectively. These results could imply that earnings announcements have increased in importance as a source of information. Table 11 also reports statistics for rms with and without December scal year-ends. Consistent with the results of Bamber, Christensen, and Gaver [2000], the mean (median) abnormal volume at the four earnings releases is 1.55% (0.89%) of the annual volume for December scal year-end rms, compared with 2.26% (1.56%) for non-December scal year-end rms.
8. Conclusions and Some Implications

While there is ample reason and evidence to believe that earnings are a source of some but not substantial new information, no study to our knowledge has quantied the relative importance of earnings announcements in the total information environment. To do so, we estimate the R 2 from a regression of securities calendar-year returns on their four earningsannouncement window returns to measure relative informativeness. This measure allows the full reaction to earnings news to occur outside the announcement window, and hence does not assume market efciency. Our principal result is that the average quarterly earnings announcement is associated with abnormal price volatility of only 1% to 2%, approximately, of total annual volatility. We also report a sharp increase during recent years in the proportion of annual information released in the earnings event windows. The increase occurs perhaps as far back as 2000, and is particularly acute in the last three years of our sample, 2004 to 2006. It could be due to increased nancial reporting quality subsequent to Enron/Sarbanes-Oxley, a reduction in analyst forecast activity, Regulation FD, chance, product or factor market conditions, or a combination of factors. It does not appear to be caused by a change in sample composition. An increase in management forecasts released concurrently with earnings (Anilowski, Feng, and Skinner [2007]) explains only part of the change. Information arrival is slightly lower than normal both before and after earnings announcements, which is somewhat surprising. In the weeks prior
1012
to earnings announcements, one might expect information production by managers, analysts, and investors (e.g., Kim and Verrecchia [1997]), but we nd no evidence of unusual price activity. In the weeks following earnings announcements, one might expect analysts to revise their forecasts, but the evidence suggests that the earliest postannouncement revisions consist largely of incorporating the new information in the earnings announcements, rather than producing new information. In contrast, the latest forecast revision before the earnings announcements is associated with abnormal volatility, which increases substantially in recent years. We also nd that management forecasts, which unlike earnings announcements are discretionary, occur less frequently, and are forward looking, are associated with substantial information. The proportion of annual information associated with earnings event windows appears to be a concave function of rm size, in contrast with the evidence Atiase [1985] obtained from a small sample of rms and quarters. A similar result is obtained for market-to-book ratio. There is a surprising degree of similarity in results across industries. The results we report on the amount of new information released at earnings announcements have several interesting implications. They are consistent with the hypothesis that the primary economic role of accounting earnings is not to provide new information to the share market. By inference, they point more to the use of nancial statements in periodic contract settlements, notably debt and compensation (Watts and Zimmerman [1986], Holthausen and Watts [2001], Ball, Robin, and Sadka [2008]) and in conrmation of prior information (Gigler and Hemmer [1998], Ball [2001]). Earnings, and other nancial statement variables affected by earnings such as leverage and distributable capital, affect outcomes in debt and compensation contracts. These contracts do not provide for settlement on a continuous basis, but with annual or other discrete frequencies. The shortterm timeliness of earnings, in the sense of whether the information in earnings is anticipated by the stock market during the year, is not of primary importance in these uses. 19 Similarly, because accounting earnings primarily report actual outcomes, it plays an economic role in the conrmation of prior, expectational information. Accurate reporting of actual earnings outcomes exerts an accountability discipline on managers and analysts more-timely expectational statements, such as growth prospects and earnings forecasts. To the extent that managers can be viewed as having an implicit contract with investors to be truthful in such statements, this also is a type of contract settlement role for earnings. Here too, the role of earnings is not to be timely. Its role is to increase the veracity of more timely sources of information.
19 Our research design focuses on timeliness on a daily basis, and hence is less appropriate for assessing usefulness in contracting contexts, where settlement is on an annual basis, than annual-return studies such as Basu [1997]. However, one could argue that the lack of earnings timeliness is the reason these contracts are settled only annually.
1013
The results help with interpreting calls to increase timeliness, such as Lev [1989], or to incorporate more fair value information based on managers expectations into the nancial statements. The evidence that announcement windows have contributed more to return volatility in recent years could indicate that earnings has increased in importance as a source of new information. Our ndings that earnings are largely anticipated potentially explain the generally low magnitudes of analysts forecast errors. The evidence suggests that analyst forecasts issued before earnings announcements are associated with new information to the market, but immediate postannouncement forecasts are not. These results help with interpreting the economic role of analysts forecasts. Is the primary economic role of analysts to produce new information, or is it to distill the implications for expected earnings of the information that is already incorporated in prices, or both? Beaver, Lambert, and Morse [1980] address the issue of bidirectional causality between earnings and returns, an issue that is central to the Dietrich, Muller, and Riedl [2007] commentary on estimating the Basu [1997] model. Our results imply that earnings do not exert a substantial causal inuence on annual returns. The results we report on relative price volatility associated with earnings announcements also help interpret the source and magnitude of the earnings announcement premium reported in Chari, Jagannathan, and Ofer [1988], Ball and Kothari [1991], Lamont [1998], Cohen et al. [2007], and Lamont and Frazzini [2007]. This research design is simple, but we believe it sheds much light on the issue of the amount of new information conveyed in earnings and thus indirectlyon the role accounting earnings plays in the economy.
REFERENCES ANILOWSKI, C.; M. FENG; AND D. J. SKINNER. Does Earnings Guidance Affect Market Returns? The Nature and Information Content of Aggregate Earnings Guidance. Journal of Accounting & Economics 44 (2007): 3663. ATIASE, R. Predisclosure Information, Firm Capitalization and Security Price Behavior Around Earnings Announcements. Journal of Accounting Research 23 (1985): 2136. BAKER, M., AND J. C. STEIN. Market Liquidity as a Sentiment Indicator. Journal of Financial Markets 7 (2004): 27199. BALL, R. Anomalies in Relationships Between Securities Yields and YieldSurrogates. Journal of Financial Economics 6 (1978): 10326. BALL, R. Infrastructure Requirements for an Economically Efcient System of Public Financial Reporting and Disclosure. Brookings-Wharton Papers on Financial Services (2001): 12769. BALL, R., AND P. BROWN. An Empirical Evaluation of Accounting Income Numbers. Journal of Accounting Research 6 (1968): 15978. BALL, R., AND S. P. KOTHARI. Security Returns Around Earnings Announcements. The Accounting Review 66 (1991): 71838. BALL, R.; A. ROBIN; AND G. SADKA. Is Financial Reporting Shaped by Equity Markets or by Debt Markets? An International Study of Timeliness and Conservatism. Review of Accounting Studies 13 (2008): 168205.
1014
BAMBER, L. S. The Information Content of Annual Earnings Releases: A Trading Volume Approach. Journal of Accounting Research 24 (1986): 4056. BAMBER, L. S. Unexpected Earnings, Firm Size, and Trading Volume Around Quarterly Earnings Announcements. The Accounting Review 62 (1987): 51032. BAMBER, L. S.; O. BARRON; AND T. STOBER. Differential Interpretations and Trading Volume. Journal of Financial and Quantitative Analysis 34 (1999): 36986. BAMBER, L. S.; T. E. CHRISTENSEN; AND K. M. GAVER. Do We Really Know What We Think We Know? A Case Study of Seminal Research and Its Subsequent Overgeneralization. Accounting Organizations and Society 25 (2000): 10329. BASU, S. The Conservatism Principle and the Asymmetric Timeliness of Earnings. Journal of Accounting & Economics 24 (1997): 337. BEAVER, W. H. The Information Content of Annual Earnings Announcements. Journal of Accounting Research 6 (Supplement 1968): 6792. BEAVER, W. H. Financial Reporting: An Accounting Revolution, Third edition. Upper Saddle River, NJ: Prentice Hall, 1998. BEAVER, W. H.; R. LAMBERT; AND D. MORSE. The Information Content of Security Prices. Journal of Accounting & Economics 2 (1980): 328. BEEKES, W., AND P. BROWN. Do Better Governed Australian Firms Make More Informative Disclosures? Journal of Business, Finance & Accounting 33 (2006a): 42250. BEEKES, W., AND P. BROWN. On the Timeliness of Price Discovery. Working paper, Lancaster University Management School, 2006b. BERKMAN, H., AND C. TRUONG. Event Day 0? After-Hours Earnings Announcements. Journal of Accounting Research (2008): Forthcoming. BERNARD, V., AND J. THOMAS. Post-Earnings Announcement Drift: Delayed Price Response or Risk Premium? Journal of Accounting Research 27 (1989): 136. BERNARD, V., AND J. THOMAS. Evidence That Stock Prices Do Not Fully Reect the Implications of Current Earnings for Future Earnings. Journal of Accounting & Economics 13 (1990): 305 40. BOTOSAN, C. A. Disclosure Level and the Cost of Equity Capital. Accounting Review 72 (1997): 32349. BUTLER, M.; A. KRAFT; AND I. S. WEISS. The Effect of Reporting Frequency on the Timeliness of Earnings: The Cases of Voluntary and Mandatory Interim Reports. Journal of Accounting & Economics 43 (2007): 181217. CHAMBERS, A., AND S. PENMAN. Timeliness of Reporting and the Stock Price Reaction to Earnings Announcements. Journal of Accounting Research 22 (1984): 2147. CHARI, V. V.; R. JAGANNATHAN; AND A. R. OFER. Seasonalities in Security Returns: The Case of Earnings Announcements. Journal of Financial Economics 21 (1988): 10121. COHEN, D. A.; A. DEY; T. Z. LYS; AND S. V. SUNDER. Earnings Announcement Premia and the Limits to Arbitrage. Journal of Accounting & Economics 43 (2007): 15380. COLLINS, D. W.; S. P. KOTHARI; AND J. D. RAYBURN. Firm Size and the Information Content of Prices With Respect to Earnings. Journal of Accounting & Economics 9 (1987): 11138. COLLINS, D. W.; S. P. KOTHARI; J. SHANKEN; AND R. G. SLOAN. Lack of Timeliness and Noise as Explanations for the Low Contemporaneous Return-Earnings Association. Journal of Accounting & Economics 18 (1994): 289324. COLLINS, D. W.; O. Z. LI; AND H. XIE. What Drives the Increased Informativeness of Earnings Announcements Over Time? Review of Accounting Studies (2008): Forthcoming. DECHOW, P. Accounting Earnings and Cash Flows as Measures of Firm Performance: The Role of Accounting Accruals. Journal of Accounting & Economics 18 (1994): 342. DEY, M. K., AND B. RADHAKRISHNA. Who Trades Around Earnings Announcements? Evidence From Torq Data. Journal of Business Finance & Accounting 34 (2007): 26991. DIETRICH, D.; K. MULLER; AND E. RIEDL. Asymmetric Timeliness Tests of Accounting Conservatism. Review of Accounting Studies 12 (2007): 95124. EASTON, P.; T. HARRIS; AND J. OHLSON. Aggregate Accounting Earnings Can Explain Most of Security Returns: The Case of Long Return Intervals. Journal of Accounting & Economics 15 (1992): 11942.
1015
FOSTER, G.; C. OLSEN; AND T. SHEVLIN. Earnings Releases, Anomalies, and the Behavior of Securities Returns. The Accounting Review 59 (1984): 574603. FRANCIS, J.; K. SCHIPPER; AND L. VINCENT. Earnings Announcements and Competing Information. Journal of Accounting & Economics 33 (2002): 31342. FRENCH, K. R., AND R. ROLL. Stock Return Variances: The Arrival of Information and Reaction of Traders. Journal of Financial Economics 17 (1986): 526. GIGLER, F., AND T. HEMMER. On the Frequency, Quality, and Informational Role of Mandatory Financial Reports. Journal of Accounting Research 36 (1998): 11747. HARRIS, M., AND A. RAVIV. Differences of Opinion Make a Horse Race. Review of Financial Studies 6 (1993): 473506. HOLTHAUSEN, R. W., AND R. E. VERRECCHIA. The Effect of Informedness and Consensus on Price and Volume Behavior. The Accounting Review 65 (1990): 191208. HOLTHAUSEN, R. W., AND R. L. WATTS. The Relevance of the Value-Relevance Literature for Financial Accounting Standard Setting. Journal of Accounting & Economics 31 (2001): 375. HOSKIN, R. E.; J. S. HUGHES; AND W. E. RICKS. Evidence on the Incremental Information Content of Additional Firm Disclosures Made Concurrently With Earnings. Journal of Accounting Research 24 (Supplement 1986): 132. JEGADEESH, N., AND S. TITMAN. Returns to Buying Winners and Selling Losers: Implications for Stock Market Efciency. Journal of Finance 48 (1993): 6591. KANDEL, E., AND N. D. PEARSON. Differential Interpretation of Public Signals and Trade in Speculative Markets. Journal of Political Economy 103 (1995): 83172. KARPOFF, J. The Relation Between Price Change and Trading Volume: A Survey. Journal of Financial and Quantitative Analysis 22 (1987): 10926. KIM, O., AND R. E. VERRECCHIA. Trading Volume and Price Reactions to Public Announcements. Journal of Accounting Research 29 (1991a): 30221. KIM, O., AND R. E. VERRECCHIA. Market Reactions to Anticipated Announcements. Journal of Financial Economics 30 (1991b): 273309. KIM, O., AND R. E. VERRECCHIA. Market Liquidity and Volume Around Earnings Announcements. Journal of Accounting & Economics 17 (1994): 4167. KIM, O., AND R. E. VERRECCHIA. Pre-announcement and Event-Period Private Information. Journal of Accounting & Economics 24 (1997): 395419. KOTHARI, S. P., AND R. G. SLOAN. Information in Prices About Future Earnings. Journal of Accounting & Economics 15 (1992): 14371. LAMONT, O. A. Earnings and Expected Returns. Journal of Finance 53 (1998): 156387. LAMONT, O. A., AND A. FRAZZINI. The Earnings Announcement Premium and Trading Volume. Working paper, National Bureau of Economic Research, 2007. LANDSMAN, W., AND E. MAYDEW. Has the Information Content of Quarterly Earnings Announcements Declined in the Past Three Decades? Journal of Accounting Research 40 (2002): 797808. LEE, C. M. C. Earnings News and Small Traders: An Intraday Analysis. Journal of Accounting & Economics 15 (1992): 265302. LEV, B. On the Usefulness of Earnings and Earnings Research: Lessons and Directions From Two Decades of Empirical Research. Journal of Accounting Research 27 (Supplement 1989): 15392. MORSE, D. Price and Trading Volume Reaction Surrounding Earnings Announcements: A Closer Examination. Journal of Accounting Research 19 (1981): 37483. NICHOLS, D., AND J. WAHLEN. How Do Earnings Numbers Relate to Stock Returns? A Review of Classic Accounting Research With Updated Evidence. Accounting Horizons 18 (2004): 26386. ODEAN, T. Do Investors Trade Too Much? American Economic Review 89 (1999): 127998. OPPONG, A. Information Content of Annual Earnings Announcements Revisited. Journal of Accounting Research 18 (1980): 57484. PENMAN, S. Insider Trading and the Dissemination of Firms Forecast Information. Journal of Business 55 (1982): 479503.
1016
ROGERS, J. L.; D. J. SKINNER; AND A. VAN BUSKIRK. Earnings Guidance and Market Uncertainty. Working paper, University of Chicago, 2008. SKINNER, D. J. Earnings Disclosures and Stockholder Lawsuits. Journal of Accounting & Economics 23 (1997): 24982. SMITH, M. J. Evidence of a Recent Increase in the Usefulness of Quarterly Earnings Announcements. Working paper, Boston University, September 2007. WATTS, R., AND J. ZIMMERMAN. Positive Accounting Theory. Englewood Cliffs, NJ: Prentice-Hall, 1986. WAYMIRE, G. Earnings Volatility and Voluntary Management Forecast Disclosure. Journal of Accounting Research 23 (1985): 26895.

34850173

Загружено:

Сведения о документе

Исходное описание:

Авторское право

Доступные форматы

Поделиться этим документом

Поделиться или встроить документ

Параметры публикации

Этот документ был вам полезен?

Это неприемлемый материал?

Авторское право:

Доступные форматы

34850173

Загружено:

Авторское право:

Доступные форматы

DOI: 10.1111/j.1475-679X.2008.00299.x Journal of Accounting Research Vol. 46 No. 5 December 2008 Printed in U.S.A.

How Much New Information Is There in Earnings?

, University of Chicago on behalf of the Institute of Professional Accounting, 2008

R. BALL AND L. SHIVAKUMAR

NEW INFORMATION IN EARNINGS

R. BALL AND L. SHIVAKUMAR

NEW INFORMATION IN EARNINGS

R. BALL AND L. SHIVAKUMAR

NEW INFORMATION IN EARNINGS

R. BALL AND L. SHIVAKUMAR

3. Research Design and Sample

NEW INFORMATION IN EARNINGS

BENCHMARKING THE REGRESSION

R. BALL AND L. SHIVAKUMAR

4. Event-Window Contribution to Annual Return Volatility

NEW INFORMATION IN EARNINGS

R. BALL AND L. SHIVAKUMAR

R 2 : NEW INFORMATION IN EARNINGS

NEW INFORMATION IN EARNINGS

1.113 0.20 1.120 0.00 1.085 1.143

5.9 6.1 6.0 5.8

3,682 3,385 2,492 4,941

R. BALL AND L. SHIVAKUMAR

1,307 1,368 1,005 1,628

NEW INFORMATION IN EARNINGS

1.036 0.47 1.045 0.32 1.033 1.038

9.3 9.0 8.6 9.9

3,682 3,385 2,492 4,941

R. BALL AND L. SHIVAKUMAR

1,307 1,368 1,005 1,628

DELETING EXTREME OBSERVATIONS

DECEMBER FISCAL YEAR-END FIRMS

NEW INFORMATION IN EARNINGS

GOOD AND BAD AGGREGATE NEWS

SLOPE COEFFICIENTS: MISPRICING EARNINGS

R. BALL AND L. SHIVAKUMAR

INDIVIDUAL YEAR AND SUBPERIOD RESULTS

NEW INFORMATION IN EARNINGS

R. BALL AND L. SHIVAKUMAR

CONCURRENT MANAGEMENT FORECASTS

NEW INFORMATION IN EARNINGS

R. BALL AND L. SHIVAKUMAR

NEW INFORMATION IN EARNINGS

R. BALL AND L. SHIVAKUMAR

NEW INFORMATION IN EARNINGS

R. BALL AND L. SHIVAKUMAR

NEW INFORMATION IN EARNINGS

R. BALL AND L. SHIVAKUMAR

EFFECT OF THE MARKET-TO-BOOK RATIO, LEVERAGE, AND INDUSTRY

NEW INFORMATION IN EARNINGS

DIFFERENT EVENT WINDOWS

5. Information in Adjacent Nonevent Windows

R. BALL AND L. SHIVAKUMAR

NEW INFORMATION IN EARNINGS

6. Analyst Earnings Forecasts and Management Forecasts

EARLIEST POSTANNOUNCEMENT ANALYST FORECAST REVISION

R. BALL AND L. SHIVAKUMAR

NEW INFORMATION IN EARNINGS

477 369 686 269

LATEST PREANNOUNCEMENT ANALYST FORECAST REVISION

R. BALL AND L. SHIVAKUMAR

513 381 671 296

513 381 671 296